This course studies large cardinals as a systematic extension of the axioms of set theory beyond ZFC. The central idea is that some infinite cardinals are so large that the universe of sets reflects deep structural features around them: they carry ultrafilters, generate elementary embeddings, satisfy strong reflection principles, or support regularity phenomena far beyond what ZFC alone can prove. The course treats large cardinals both as mathematical objects and as a measuring scale for the strength of axioms. The chapters begin with the large cardinal hierarchy and the reflection principles that motivate it, then move to the first major dividing line: measurable cardinals, ultrafilters, ultrapowers, and elementary embeddings. From there the course develops stronger embedding axioms, introduces Woodin cardinals, and studies the tension between large cardinals and the constructible universe $L$. The later chapters turn toward inner model theory, determinacy, descriptive set theory, and regularity properties, showing how large cardinals influence both the structure of the universe and the behavior of definable sets of reals. By the end, the course builds a calibrated picture of consistency strength: which principles imply which others, which theories rule out certain canonical universes, and how modern set theory compares axioms by their consequences. The capstone chapter ties together hierarchy, reflection, embeddings, inner models, and determinacy into the contemporary view of large cardinals as a central organizing framework for the foundations of mathematics. # Introduction Large cardinal axioms extend ZFC by asserting the existence of cardinals whose size is reflected in structural, model-theoretic, or combinatorial properties. The course studies these axioms as a calibrated hierarchy rather than as isolated principles: inaccessible cardinals begin the hierarchy by making initial segments $V_\kappa$ resemble the universe, measurable cardinals introduce elementary embeddings and ultrafilters, and Woodin cardinals connect the hierarchy to determinacy and inner model theory. The guiding question is how set theory measures the strength of assertions it cannot settle from ZFC alone. Instead of using forcing as the central technique, the course uses large cardinals to organize relative consistency, reflection, elementary embeddings, and canonical inner models. The intended background is a first graduate course in set theory: ordinals, cardinals, transfinite recursion, stationary sets, filters, ultrafilters, basic model theory, $L$, descriptive set theory, and the role of forcing in independence. Later chapters also quote a small number of advanced inner-model and determinacy theorems as landmarks; those results are used for orientation and calibration, not proved from first principles in these notes. ## Why Large Cardinals Enter Set Theory What kind of mathematical evidence can support axioms that go beyond ZFC? The first answer is closure: some cardinals are so large that the cumulative hierarchy below them satisfies substantial fragments, and sometimes all, of the ordinary axioms of set theory. The second answer is reflection: properties of the universe often appear already inside smaller ranks. The third answer is comparison of strength: if one principle gives models of another, then it occupies a higher place in the consistency hierarchy. [definition: Large Cardinal Axiom] A large cardinal axiom is a set-theoretic assertion extending ZFC by stating that there exists a cardinal $\kappa$ with a specified strong closure, reflection, compactness, ultrafilter, or elementary-embedding property. [/definition] This definition is deliberately schematic, because the course will encounter several incompatible-looking formulations that become part of one hierarchy. The same cardinal may be described combinatorially, through filters, through elementary embeddings, or through reflection principles; comparing these descriptions is one of the main tasks of the subject. [example: First Large Cardinal Themes] An inaccessible cardinal packages two closure requirements into one cardinal property: $\kappa$ is uncountable and regular, so every sequence of length $<\kappa$ whose entries are ordinals below $\kappa$ has supremum below $\kappa$; and $\kappa$ is a strong limit, so for every $\lambda<\kappa$ one has $2^\lambda<\kappa$. These two clauses say different things. Regularity controls the length of constructions, while the strong limit condition controls the size of power sets formed below $\kappa$. A measurable cardinal is presented in a different language. Instead of first asking that $V_\kappa$ be closed under set-forming operations, one asks for a non-principal $\kappa$-complete ultrafilter $U$ on $\kappa$. Non-principality means no single point $\alpha<\kappa$ generates the measure: $\{\alpha\}\notin U$ for every $\alpha<\kappa$. $\kappa$-completeness means that whenever $\lambda<\kappa$ and $A_i\in U$ for each $i<\lambda$, the intersection $\bigcap_{i<\lambda}A_i$ is still in $U$. Later, this same structure is reexpressed by forming the ultrapower map $j:V\to\operatorname{Ult}(V,U)$, where $j(x)$ is the equivalence class of the constant function with value $x$, and the first ordinal moved by $j$ is $\kappa$. A Woodin cardinal is described in still another language: instead of one measure on $\kappa$, one uses a coherent supply of extenders or elementary embeddings that reflect information about subsets of $V_\delta$. The point of comparing these examples is that “large cardinal” is not one definition repeated at larger sizes. It is a family of principles in which closure, measures, embeddings, and reflection become stronger ways for the universe to see its own initial segments. [/example] The next issue is methodological. ZFC cannot prove that these cardinals exist, and in many cases ZFC cannot prove their nonexistence without collapsing its own consistency. The subject therefore treats implication between consistency statements as a mathematical measuring device. [definition: Relative Consistency Strength] Let $T$ and $S$ be theories extending a fixed base theory such as ZFC. We say that $T$ has at least the relative consistency strength of $S$ if, over the accepted metatheory, $\operatorname{Con}(T)$ implies $\operatorname{Con}(S)$. [/definition] This definition records only consistency implication, not interpretability, proof-theoretic reducibility, or philosophical priority. In practice, large cardinal theory often proves stronger statements: from a model of one theory, it builds an inner model, an ultrapower, or a rank-initial segment satisfying another theory. [example: Inaccessibles and Models of ZFC] Assume $\kappa$ is an uncountable inaccessible cardinal, so $\kappa$ is regular and strong limit. The rank-initial structure $(V_\kappa,\in)$ satisfies ZFC by the closure verification below. Since $V_\kappa$ is a set, this gives a set model of ZFC in the ambient metatheory: take $M=V_\kappa$ and $E=\in\cap(V_\kappa\times V_\kappa)$, so $(M,E)\models\operatorname{ZFC}$. The closure points in the verification are exactly the two parts of inaccessibility. If $x\in V_\kappa$, then there is some $\alpha<\kappa$ with $x\in V_\alpha$, hence $x\subseteq V_\alpha$ and every subset of $x$ has rank $<\alpha+1<\kappa$; therefore $\mathcal P(x)\in V_\kappa$. For Replacement, if $a\in V_\kappa$ and a definable function sends each $u\in a$ to some $F(u)\in V_\kappa$, then for each $u\in a$ choose $\beta_u<\kappa$ with $F(u)\in V_{\beta_u}$. Because $a\in V_\kappa$, the set of indices $\{\beta_u:u\in a\}$ has size $<\kappa$, and regularity gives \begin{align*} \sup\{\beta_u:u\in a\}<\kappa. \end{align*} Thus the whole image $\{F(u):u\in a\}$ is contained in some $V_\gamma$ with $\gamma<\kappa$, so it is itself an element of $V_\kappa$. Consequently, in the ambient metatheory, \begin{align*} \operatorname{ZFC}+\text{``there is an inaccessible cardinal''}\vdash \exists M\,((M,\in)\models\operatorname{ZFC}). \end{align*} A set model of ZFC yields $\operatorname{Con}(\operatorname{ZFC})$, so the inaccessible-cardinal assumption has at least the relative consistency strength of ZFC itself. This is the first model-theoretic calibration in the course: one large cardinal produces a universe-like initial segment. [/example] ## The Hierarchy as a Measuring Device How can cardinals be compared when their definitions involve different mathematical languages? Some definitions speak about the size of power sets, some about trees, some about ultrafilters, and some about elementary embeddings. The hierarchy becomes meaningful because these notions imply one another, produce models of one another, and often line up with independent combinatorial or descriptive-set-theoretic consequences. [definition: Consistency Hierarchy] A consistency hierarchy is an ordering of theories by relative consistency strength, where $T$ is placed above $S$ when $\operatorname{Con}(T)$ entails $\operatorname{Con}(S)$ and no converse implication is known from the accepted base theory. [/definition] After the definition, the point to remember is that the hierarchy is not merely a list of larger and larger cardinals. To see why the first entry deserves its place, we need a concrete theorem showing that a cardinal property gives an actual model of ZFC, not just an impressive closure condition. [quotetheorem:7397] [citeproof:7397] This theorem explains why inaccessibility is the first natural stopping point in the cumulative hierarchy, and it also shows why both hypotheses matter. If $\kappa$ is a limit cardinal but not a strong limit, there may be some $\alpha<\kappa$ with $|V_\alpha|\geq\kappa$; then the cardinal-bound part of the Replacement argument breaks down, because a set already in $V_\kappa$ can have enough elements to support images cofinal in $\kappa$. If $\kappa$ is a singular strong limit, Power Set behaves well but Replacement can fail: a cofinal sequence $(\alpha_i)_{i<\operatorname{cf}(\kappa)}$ in $\kappa$ is indexed by a set in $V_\kappa$, while its image is unbounded in rank below $\kappa$ and therefore need not be collected into a single set of $V_\kappa$. Thus the proof is not a formal slogan that large ranks satisfy ZFC; it is a checklist where strong limit supplies cardinal closure and regularity controls images of definable functions. This also previews the course's repeated pattern for reading consistency implications. A large cardinal assumption is useful when it produces a concrete model, ultrapower, or inner model satisfying a target theory; the relative consistency statement is then read off from that construction in the metatheory. [example: Grothendieck-Style Universes] Let $\kappa$ be inaccessible and put $U=V_\kappa$. We verify the usual universe-style closure properties by bounding ranks below $\kappa$. If $x,y\in U$, choose $\alpha,\beta<\kappa$ with $x\in V_\alpha$ and $y\in V_\beta$. Let $\gamma=\max\{\alpha,\beta\}+1$. Since $\kappa$ is a limit ordinal, $\gamma+1<\kappa$, and $x,y\in V_\gamma$, so \begin{align*} \{x,y\}\subseteq V_\gamma. \end{align*} By the definition of the cumulative hierarchy, every subset of $V_\gamma$ belongs to $V_{\gamma+1}$, hence $\{x,y\}\in V_{\gamma+1}\subseteq V_\kappa$. If $x\in U$, choose $\alpha<\kappa$ with $x\in V_\alpha$. Then every element of $x$ has rank below $\alpha$, so \begin{align*} x\subseteq V_\alpha. \end{align*} Therefore every subset of $x$ is also a subset of $V_\alpha$, and hence \begin{align*} \mathcal P(x)\subseteq \mathcal P(V_\alpha)=V_{\alpha+1}. \end{align*} Thus $\mathcal P(x)\in V_{\alpha+2}\subseteq V_\kappa$, because $\alpha+2<\kappa$. Now let $(x_i)_{i\in I}$ be an indexed family coded by a function $f\in U$ with domain $I\in U$ and $f(i)=x_i$ for each $i\in I$. Choose $\alpha<\kappa$ with $f\in V_\alpha$. For each $i\in I$, the value $x_i$ occurs inside an ordered pair $(i,x_i)\in f$, so its rank is bounded below $\alpha$; hence $x_i\in V_\alpha$ for every $i\in I$. The union of the family is therefore bounded by \begin{align*} \bigcup_{i\in I}x_i\subseteq V_\alpha. \end{align*} It follows that $\bigcup_{i\in I}x_i\in V_{\alpha+1}\subseteq V_\kappa$. Thus $V_\kappa$ behaves like a universe closed under the set-forming operations commonly used in ordinary mathematics, which is why inaccessible cardinals naturally appear in category-theoretic and foundational settings. [/example] ## The Elementary-Embedding Viewpoint What changes when large cardinals are no longer described only by closure properties of $V_\kappa$? The subject becomes much more powerful once it studies elementary maps into transitive target models. A non-identity global elementary embedding $j:V\to M$ is not a set inside $V$, and in ordinary ZFC/NBG it must be treated with care: unrestricted global embedding hypotheses run into Kunen-style inconsistency phenomena. The embedding language used in the course is therefore supplied by precise constructions, such as well-founded ultrapowers, extenders, or rank-initial approximations with enough domain to discuss the relevant subsets. [definition: Critical Point] Let $j:M\to N$ be an elementary embedding between transitive classes or transitive sets. The critical point of $j$ is the least ordinal $\kappa$ such that $j(\kappa)\ne \kappa$. [/definition] The critical point is where the embedding first detects large-cardinal strength. Below it, the embedding is fixed on all ordinals; at it, the embedding moves the hierarchy upward and creates structural information about subsets, measures, and reflection. [example: Ultrapower Intuition] Let $U$ be a $\kappa$-complete non-principal ultrafilter on $\kappa$. Define $f\sim_U g$ for functions $f,g:\kappa\to V$ exactly when $\{\xi<\kappa:f(\xi)=g(\xi)\}\in U$, and write $[f]_U$ for the equivalence class of $f$. The ultrapower map sends each set $x$ to $j(x)=[c_x]_U$, where $c_x(\xi)=x$ for every $\xi<\kappa$. We verify why the first ordinal moved is $\kappa$. If $B\subseteq\kappa$ has size $<\kappa$, then $B\notin U$: for each $\beta\in B$, non-principality gives $\{\beta\}\notin U$, so $\kappa\setminus\{\beta\}\in U$, and $\kappa$-completeness gives \begin{align*} \bigcap_{\beta\in B}(\kappa\setminus\{\beta\})=\kappa\setminus B\in U. \end{align*} Thus every bounded subset of $\kappa$ is outside $U$, and every tail $\{\xi<\kappa:\alpha<\xi\}$ is in $U$. For each $\alpha<\kappa$, the constant class $[c_\alpha]_U$ represents the same ordinal $\alpha$ in the ultrapower, by induction on $\alpha$: if all smaller ordinals are fixed, then \begin{align*} j(\alpha)=j(\{\beta:\beta<\alpha\})=\{j(\beta):\beta<\alpha\}=\{\beta:\beta<\alpha\}=\alpha. \end{align*} Let $\operatorname{id}(\xi)=\xi$. For every $\alpha<\kappa$, \begin{align*} \{\xi<\kappa:c_\alpha(\xi)<\operatorname{id}(\xi)\}=\{\xi<\kappa:\alpha<\xi\}\in U, \end{align*} so $\alpha<[ \operatorname{id}]_U$ in the ultrapower. Also, \begin{align*} \{\xi<\kappa:\operatorname{id}(\xi)<c_\kappa(\xi)\}=\{\xi<\kappa:\xi<\kappa\}=\kappa\in U, \end{align*} so $[\operatorname{id}]_U<j(\kappa)=[c_\kappa]_U$. Hence every ordinal below $\kappa$ is fixed, while $j(\kappa)>\kappa$. This is the sense in which the ultrafilter presentation of measurability produces the elementary-embedding viewpoint: the measure determines an ultrapower embedding whose critical point is exactly $\kappa$. [/example] The ultrapower example raises the reverse question that will recur throughout the course: if enough of an elementary embedding is given first, what large-cardinal structure can be recovered from it? The next theorem states the safe local form of the extraction argument. It uses only a domain containing the subsets of $\kappa$ and their short sequences, rather than an unrestricted global embedding of the whole universe. [quotetheorem:7398] [citeproof:7398] This theorem is one of the first bridges between two languages: sufficiently rich embeddings yield measures, but every hypothesis prevents a specific failure. Transitivity of $N$ lets membership and ordinal comparison be read correctly in the target. The requirement that all subsets of $\kappa$ lie in the domain makes $U$ a measure on the full $\mathcal P(\kappa)$, not merely on a fragment. The closure under short coded sequences is what turns the finite ultrafilter argument into $\kappa$-completeness. The restriction to a local or construction-derived embedding avoids the dangerous global assumption of a free-standing non-identity $j:V\to M$, which is not an available hypothesis in ordinary ZFC/NBG and is constrained by Kunen-style inconsistency theorems. This direction should be distinguished from the ultrapower construction. Starting with a measure, one builds an embedding by quotienting functions modulo the ultrafilter and then proving well-foundedness or taking a transitive collapse. Starting with an embedding, the measure is recovered by the single test $A\in U$ iff $\kappa\in j(A)$. Later chapters reverse and generalize these procedures through ultrapowers and extenders. ## Reflection, Inner Models, and Determinacy Why do large cardinal axioms appear far from their original definitions, for instance in descriptive set theory or canonical inner models? Strong reflection principles tend to impose regularity on definable sets of reals, while inner model theory tries to build canonical universes containing enough large-cardinal structure to explain that regularity. The course uses these links as applications of the hierarchy rather than as separate subjects. [definition: Inner Model] An inner model is a transitive proper class $M$ such that $M$ contains all ordinals and $(M,\in)$ satisfies ZFC. [/definition] Inner models let us compare universes without leaving the language of membership. The constructible universe $L$ is the first example; more advanced inner models incorporate measurable, strong, or Woodin cardinals through fine structure and extender sequences. [example: The Role of $L$] The constructible universe $L$ is defined by transfinite recursion: $L_0=\varnothing$, $L_{\alpha+1}=\operatorname{Def}(L_\alpha)$, $L_\lambda=\bigcup_{\alpha<\lambda}L_\alpha$ at limit stages, and $L=\bigcup_{\alpha\in\operatorname{Ord}}L_\alpha$. This construction gives a canonical well-ordering by ordering each new element of $L_{\alpha+1}$ according to the first formula and parameters from $L_\alpha$ that define it, and then ordering elements first by the least stage at which they appear. With this well-ordering, $L$ satisfies ZFC and is the basic inner model used as a comparison point. Here $0^\#$ denotes the sharp for $L$. One way to read its existence is: there is a nontrivial class of indiscernibles for $L$, meaning a class $I$ of ordinals such that whenever $\alpha_1<\cdots<\alpha_n$ and $\beta_1<\cdots<\beta_n$ are increasing sequences from $I$, the structure $L$ satisfies the same formulas with parameters $\alpha_1,\ldots,\alpha_n$ as with parameters $\beta_1,\ldots,\beta_n$. Thus the truth pattern of $L$ cannot distinguish between different increasing finite blocks of indiscernibles. If $0^\#$ exists in $V$, then these indiscernibles are present from the viewpoint of $V$ but are not part of the internal construction of $L$ itself. So $L$ remains a canonical model of ZFC with a definable global well-order, while $V$ contains additional structure showing that $L$ has missed some large-cardinal strength. This is why $L$ is both a model to study and a baseline to move beyond: many large cardinal hypotheses assert that the ambient universe has structure invisible to the constructible universe. [/example] The example of $L$ shows that inner models measure how much structure is visible inside a canonical subuniverse. To connect this with descriptive set theory, we need a second kind of organizing principle: games whose winning strategies encode regularity properties of sets of reals. [definition: Determinacy Principle] A determinacy principle asserts that every game in a specified class of infinite two-player perfect-information games is determined, meaning that one of the players has a winning strategy. [/definition] The definition does not specify the class of games, because different choices give very different strengths. Borel determinacy is provable in ZFC, projective determinacy requires substantial large-cardinal strength, and stronger determinacy axioms move beyond Choice in settings such as $L(\mathbb R)$. [example: Projective Regularity] Assume a large-cardinal hypothesis strong enough to imply projective determinacy, for example the usual Woodin-cardinal hypotheses used later in the course. Let $A\subseteq\mathbb R$ be projective. Projective determinacy says that every projective game with payoff set projectively definable from the same parameters as $A$ is determined: one of the two players has a strategy that wins every play compatible with it. The determinacy-to-regularity step then applies separately to the three regularity properties. For Lebesgue measurability, one associates to $A$ a game whose winning strategies code approximations of $A$ by Borel sets modulo null sets; determinacy gives a winning strategy, and the strategy determines a Borel set $B$ such that $A\triangle B$ is null. For the Baire property, the analogous game codes approximation by open sets modulo meager sets; the winning strategy gives an [open set](/page/Open%20Set) $O$ such that $A\triangle O$ is meager. For the perfect set property, the game is arranged so that either a winning strategy for one player codes a countable enumeration of $A$, or a winning strategy for the other player builds a perfect tree $T\subseteq 2^{<\omega}$ whose set of branches satisfies $[T]\subseteq A$. Thus the large-cardinal assumption is not used directly to measure subsets of the real line. It first produces determinacy for the relevant projective games, and the winning strategies supplied by determinacy are then converted into the concrete regularity witnesses: a Borel approximation for measurability, an open approximation for category, and either countability or a perfect subset for the perfect set property. [/example] ## How These Notes Are Organized What should the reader track from chapter to chapter? The first theme is the cardinal property itself: closure, compactness, measurability, strength, supercompactness, or Woodinness. The second theme is the mathematical language used to express it: stationary reflection, trees, ultrafilters, elementary embeddings, extenders, or games. The third theme is the consistency comparison produced by that property. [explanation: Course Roadmap] The course begins with inaccessible and Mahlo cardinals, where the hierarchy can still be understood through closure, regularity, and stationary sets. It then moves to reflection and indescribability, including weak compactness and tree properties. Measurable cardinals introduce countably complete ultrafilters and ultrapower embeddings, after which stronger cardinals are developed using the elementary-embedding viewpoint. Chapters 6 through 10 explain how Woodin cardinals interact with inner model theory, determinacy, projective regularity, and generic absoluteness, giving a bridge from higher infinity to definable sets of reals. [/explanation] The course will often state results in a form that emphasizes method rather than maximum generality. When a theorem is used as a measuring device, the notes will identify the hypothesis, the model or embedding it produces, and the consistency consequence. When a result is quoted from deeper inner model theory or descriptive set theory, the notes will say so and use it as context rather than hiding the extra machinery. [remark: Background Use of Forcing] Forcing is assumed as background for understanding independence and relative consistency, but it is not the main technical engine of these notes. The course uses forcing terminology when it clarifies why a statement is independent or why a consistency implication matters. The construction of forcing extensions is not redeveloped here. [/remark] The chapter therefore sets the vocabulary for the rest of the notes. Large cardinals will be treated simultaneously as axioms of infinity, as sources of elementary embeddings, as generators of inner models, and as calibrators for the strength of mathematical statements beyond ZFC. With the basic roles of large cardinals now in place, the next step is to organize them into a hierarchy. We begin at the point where largeness means that $V_\kappa$ itself reflects enough of the universe to serve as a miniature model of set theory. # 1. The Large Cardinal Hierarchy Large cardinal axioms begin with a shift in viewpoint: instead of asking how large a particular set is, we ask when an initial segment $V_\kappa$ of the universe behaves like a universe of set theory in its own right. This chapter develops the first levels of that hierarchy. The main theme is that closure properties of a cardinal $\kappa$ become model-theoretic strength when they are strong enough to make $V_\kappa$ satisfy the axioms of ZFC. The prerequisites are the cumulative hierarchy, cofinality and regular cardinals, basic cardinal arithmetic, and the model-theoretic notation $M\models T$. We also use the club and stationary subsets of a regular uncountable cardinal, recalled below before Mahloness is introduced. The hierarchy starts with inaccessibility, then strengthens it through reflection properties such as Mahloness. These cardinals are not introduced as isolated curiosities: they are the measuring rods used later for comparing measurable, strong, supercompact, and Woodin cardinals. The chapter also fixes the language of relative consistency, since large cardinal theory studies axioms whose consistency cannot be proved from ZFC itself if ZFC is consistent. ## Regular Strong Limit Cardinals and Inaccessibility The first problem is to identify which cardinals are too large to be reached by the ordinary operations of set formation below them. Taking successors, forming power sets, and taking unions of shorter sequences are the basic ways ZFC builds larger sets. A cardinal that is closed under these operations is a natural candidate for supporting a self-contained rank-initial universe. [definition: Regular Cardinal] A cardinal $\kappa$ is regular if $\operatorname{cf}(\kappa)=\kappa$. [/definition] Regularity says that $\kappa$ cannot be assembled as the supremum of fewer than $\kappa$ smaller ordinals. The key example is $\omega$, while $\aleph_\omega$ is singular because it is the supremum of the sequence $(\aleph_n)_{n\in\mathbb N}$. [example: Singular Limit Cardinal] The cardinal $\aleph_\omega$ has cofinality $\omega$. The sequence \begin{align*} \aleph_1<\aleph_2<\cdots<\aleph_n<\cdots \end{align*} is increasing and indexed by $\omega$, and by the definition of $\aleph_\omega$ as the first limit of the sequence $(\aleph_n)_{n<\omega}$ we have \begin{align*} \sup_{1\leq n<\omega}\aleph_n=\aleph_\omega. \end{align*} Thus there is a cofinal subset of $\aleph_\omega$ of order type $\omega$, so $\operatorname{cf}(\aleph_\omega)\leq\omega$. It remains to rule out finite cofinal subsets. Let $F\subset\aleph_\omega$ be finite. For each $\alpha\in F$, choose $n_\alpha<\omega$ with $\alpha<\aleph_{n_\alpha}$, which is possible because $\aleph_\omega=\sup_{n<\omega}\aleph_n$. Since $F$ is finite, the set $\{n_\alpha:\alpha\in F\}$ has a maximum $N<\omega$. Then every $\alpha\in F$ satisfies $\alpha<\aleph_N$, so \begin{align*} \sup F\leq \aleph_N<\aleph_\omega. \end{align*} Therefore no finite subset of $\aleph_\omega$ is cofinal, so $\operatorname{cf}(\aleph_\omega)\neq n$ for every finite $n$. Since $\operatorname{cf}(\aleph_\omega)\leq\omega$ and it is not finite, we get $\operatorname{cf}(\aleph_\omega)=\omega$. This shows why being a limit cardinal does not by itself give the regularity required for inaccessibility. [/example] The example isolates the failure of closure under short increasing unions. Even if this defect is removed by regularity, a second route to large sets remains: the power set of a smaller set may already have size at least $\kappa$. To make $V_\kappa$ closed under power sets of its members, we need a condition that controls $2^\lambda$ for every $\lambda<\kappa$. [definition: Strong Limit Cardinal] A cardinal $\kappa$ is a strong limit cardinal if for every cardinal $\lambda<\kappa$, \begin{align*} 2^\lambda<\kappa. \end{align*} [/definition] Strong limit cardinals are closed under many bounded cardinal arithmetic operations. For example, if $\lambda,\mu<\kappa$, then $\lambda^\mu<\kappa$ whenever $\kappa$ is a strong limit and $\mu<\kappa$, since $\lambda^\mu\leq 2^{\mu\cdot\lambda}<\kappa$ after bounding $\mu\cdot\lambda$ below $\kappa$. [definition: Strongly Inaccessible Cardinal] An uncountable cardinal $\kappa$ is strongly inaccessible if $\kappa$ is regular and a strong limit cardinal. [/definition] The word "uncountable" excludes $\omega$, which is regular and a strong limit under the finite interpretation of smaller cardinal exponentiation. In large cardinal theory, inaccessible cardinals are meant to be genuinely beyond the cumulative behaviour of the finite and countable ranks. [example: Closure of $V_\kappa$ Under Small Functions] Let $\kappa$ be strongly inaccessible, let $\lambda<\kappa$, and let $f:\lambda\to V_\kappa$. For each $\xi<\lambda$, choose an ordinal $\rho(\xi)<\kappa$ such that $f(\xi)\in V_{\rho(\xi)}$. We show that the whole graph of $f$ has rank below $\kappa$, and hence $f\in V_\kappa$. The set of ordinals \begin{align*} R=\{\rho(\xi):\xi<\lambda\} \end{align*} has cardinality at most $\lambda<\kappa$. Since $\kappa$ is regular, every subset of $\kappa$ of cardinality below $\kappa$ has supremum below $\kappa$, so \begin{align*} \beta=\sup R<\kappa. \end{align*} Thus for every $\xi<\lambda$ we have $\rho(\xi)\leq\beta$ or $\rho(\xi)<\beta+1$, and therefore \begin{align*} f(\xi)\in V_{\beta+1}. \end{align*} Also $\xi<\lambda$, so $\xi\in V_\lambda$. Let \begin{align*} \gamma=\max\{\lambda,\beta+1\}. \end{align*} Then $\gamma<\kappa$, and for every $\xi<\lambda$ both $\xi$ and $f(\xi)$ lie in $V_\gamma$. Using the Kuratowski ordered pair $\langle a,b\rangle=\{\{a\},\{a,b\}\}$, if $a,b\in V_\gamma$, then $\{a\},\{a,b\}\in V_{\gamma+1}$ and hence \begin{align*} \langle a,b\rangle\in V_{\gamma+2}. \end{align*} Therefore every pair $\langle \xi,f(\xi)\rangle$ lies in $V_{\gamma+2}$, so \begin{align*} f=\{\langle \xi,f(\xi)\rangle:\xi<\lambda\}\subseteq V_{\gamma+2}. \end{align*} It follows that \begin{align*} f\in V_{\gamma+3}. \end{align*} Since $\kappa$ is an uncountable regular cardinal, it is a limit ordinal, so $\gamma+3<\kappa$. Hence $f\in V_\kappa$. The point is that regularity prevents the ranks of fewer than $\kappa$ many values from climbing cofinally to $\kappa$, so small functions into $V_\kappa$ are themselves captured inside $V_\kappa$. [/example] This closure property explains why inaccessible cardinals have model-theoretic significance. The axioms of ZFC repeatedly demand closure under operations such as pairing, union, power set, and replacement images. The next theorem records that the two ingredients in strong inaccessibility are exactly what is needed for these operations to remain inside $V_\kappa$. [quotetheorem:7399] [citeproof:7399] Both hypotheses are doing separate work. Regularity is needed for Replacement: if $\kappa$ is singular, a sequence of length $\operatorname{cf}(\kappa)$ can have ranks cofinal in $\kappa$, so the image of a small set under a definable operation need not be captured in any smaller $V_\alpha$. The strong limit condition is needed for Power Set: if some $\lambda<\kappa$ satisfies $2^\lambda\geq\kappa$, then a set of rank below $\kappa$ and size $\lambda$ may have too many subsets for its full power set to lie inside $V_\kappa$. The theorem also has an important limitation. It says that a strongly inaccessible cardinal produces a rank-initial model of ZFC, not that ZFC proves such a cardinal exists. This gives the first sense in which large cardinals increase consistency strength: if ZFC proved that a strongly inaccessible cardinal exists, then ZFC would prove that there is a model of ZFC, contradicting [Gödel's second incompleteness theorem](/theorems/1508) provided ZFC is consistent and formalised in the usual way. [example: $V_\kappa$ as a Grothendieck Universe] Let $\kappa$ be strongly inaccessible and set $U=V_\kappa$. We verify the Grothendieck-universe closure properties from the rank definition of $V_\kappa$. If $y\in x\in U$, choose $\alpha<\kappa$ with $x\in V_\alpha$. Since every $V_\alpha$ is transitive, $y\in V_\alpha\subseteq V_\kappa$, so $U$ is transitive. If $a,b\in U$, choose $\alpha,\beta<\kappa$ with $a\in V_\alpha$ and $b\in V_\beta$, and put $\gamma=\max\{\alpha,\beta\}$. Then $a,b\in V_\gamma$, so \begin{align*} \{a,b\}\subseteq V_\gamma. \end{align*} Hence $\{a,b\}\in V_{\gamma+1}$. Since $\kappa$ is an uncountable regular cardinal, it is a limit ordinal, so $\gamma+1<\kappa$, and therefore $\{a,b\}\in U$. If $x\in U$, choose $\alpha<\kappa$ with $x\in V_\alpha$. Transitivity of $V_\alpha$ gives $x\subseteq V_\alpha$. Thus every $y\subseteq x$ satisfies $y\subseteq V_\alpha$, so \begin{align*} y\in \mathcal P(V_\alpha)=V_{\alpha+1}. \end{align*} Therefore $\mathcal P(x)\subseteq V_{\alpha+1}$, and hence \begin{align*} \mathcal P(x)\in V_{\alpha+2}\subseteq V_\kappa. \end{align*} So $U$ contains the power set of each of its members. For indexed unions, let $I\in U$ and let $\langle x_i:i\in I\rangle$ be a family with each $x_i\in U$. First $|I|<\kappa$: if $I\in V_\alpha$ for some $\alpha<\kappa$, then $|V_\alpha|<\kappa$, by induction on $\alpha$ using $|V_{\beta+1}|=2^{|V_\beta|}<\kappa$ from the strong limit property and using regularity of $\kappa$ at limit stages. For each $i\in I$, choose $\rho(i)<\kappa$ with $x_i\in V_{\rho(i)}$. Since $|I|<\kappa$ and $\kappa$ is regular, \begin{align*} \beta=\sup\{\rho(i):i\in I\}<\kappa. \end{align*} Then every $x_i$ lies in $V_{\beta+1}$, so \begin{align*} \bigcup_{i\in I}x_i\subseteq V_{\beta+1}. \end{align*} Consequently \begin{align*} \bigcup_{i\in I}x_i\in V_{\beta+2}\subseteq V_\kappa. \end{align*} Thus $V_\kappa$ is transitive and closed under pairing, power sets, and unions indexed by sets already in $V_\kappa$. These are precisely the set-forming operations that make $V_\kappa$ function as a Grothendieck universe, so the small sets relative to this universe are exactly the sets belonging to $V_\kappa$. [/example] ## Weakly Inaccessible Versus Strongly Inaccessible Cardinals The next question is how much of inaccessibility comes from cofinality alone, and how much comes from controlling the power set operation. This distinction matters because the continuum function can behave very differently in different models of set theory. Weak inaccessibility isolates the order-theoretic part. [definition: Weakly Inaccessible Cardinal] An uncountable cardinal $\kappa$ is weakly inaccessible if $\kappa$ is regular and a limit cardinal. [/definition] Every strongly inaccessible cardinal is weakly inaccessible, since a strong limit cardinal is a limit cardinal. The converse is not provable in ZFC and can fail under common cardinal arithmetic patterns. [example: Weak Inaccessibility Does Not Control Powers] Suppose $\kappa$ is an uncountable regular limit cardinal, and suppose there is a cardinal $\lambda<\kappa$ such that \begin{align*} 2^\lambda\geq\kappa. \end{align*} Because $\kappa$ is uncountable, regular, and a limit cardinal, it satisfies the definition of a weakly inaccessible cardinal. We now check why $\kappa$ is not strongly inaccessible. To be strongly inaccessible, $\kappa$ would have to be a strong limit cardinal, meaning that every cardinal $\mu<\kappa$ satisfies \begin{align*} 2^\mu<\kappa. \end{align*} Taking $\mu=\lambda$, the assumed inequality gives \begin{align*} 2^\lambda\geq\kappa. \end{align*} This is the negation of the required inequality $2^\lambda<\kappa$ for that particular smaller cardinal $\lambda$. Hence $\kappa$ is not a strong limit cardinal, and therefore it is not strongly inaccessible. Thus weak inaccessibility controls how $\kappa$ is approached by increasing sequences of smaller cardinals, through regularity and the limit-cardinal condition, but it does not by itself control the sizes of power sets below $\kappa$. [/example] The distinction collapses under the generalized continuum hypothesis, but not because the definitions become identical. Rather, GCH supplies the missing comparison between limit cardinals and power sets below them. [quotetheorem:7400] [citeproof:7400] This result is a useful calibration point because it identifies exactly what GCH supplies: it turns the order-theoretic fact that $\kappa$ is a limit cardinal into the arithmetic fact that $2^\lambda<\kappa$ for all $\lambda<\kappa$. Without GCH, the implication can fail. For instance, in models obtained by Easton-style forcing, a regular limit cardinal $\kappa$ can remain regular and a limit cardinal while the continuum function is arranged so that $2^\lambda\geq\kappa$ for some $\lambda<\kappa$; then $\kappa$ is weakly inaccessible but not strongly inaccessible. The theorem therefore does not say that weak inaccessibility by itself has the same large-cardinal strength as strong inaccessibility. It says that under a strong global hypothesis about the continuum function, the missing power-set closure follows automatically. This is why weak inaccessibility is not usually treated as a robust large cardinal notion: it is sensitive to cardinal arithmetic in a way that strong inaccessibility is designed to avoid. [remark: Terminology] Many authors write "inaccessible" for "strongly inaccessible" unless the weak/strong distinction is under discussion. In these notes, the phrase "strongly inaccessible" is used when the power set closure is part of the argument, and "weakly inaccessible" is used for the regular limit property alone. [/remark] ## Mahlo Cardinals and Stationary Reflection Inaccessibility says that $V_\kappa$ is closed under the usual set-forming operations. The next problem is to ask whether the inaccessibles below $\kappa$ occur often enough to reflect closed unbounded structure. This is the first step from being a large universe to seeing many smaller large universes inside it. We recall the club and stationary notions because Mahloness is formulated through them. The guiding idea is that a stationary set cannot be avoided by any closed unbounded approximation process. [definition: Closed Unbounded Set] Let $\kappa$ be a regular uncountable cardinal. A set $C\subset\kappa$ is closed unbounded, or club, in $\kappa$ if it is unbounded in $\kappa$ and whenever $\langle\alpha_i:i<\lambda\rangle$ is an increasing sequence from $C$ with $0<\lambda<\kappa$ and supremum $\alpha<\kappa$, then $\alpha\in C$. [/definition] The closure condition says that $C$ contains limits of its own increasing sequences of length below $\kappa$. The unboundedness condition says that $C$ reaches arbitrarily high below $\kappa$. [example: Constructing a Club by Closure Points] Let $\kappa$ be regular uncountable and let $f:\kappa\to\kappa$. We show that \begin{align*} C_f=\{\alpha<\kappa:f[\alpha]\subset\alpha\} \end{align*} is club in $\kappa$. First we prove unboundedness. Fix $\beta<\kappa$. Choose $\alpha_0$ with $\beta<\alpha_0<\kappa$. Suppose $\alpha_n<\kappa$ has been chosen. Since $|\alpha_n|<\kappa$ and $f[\alpha_n]\subseteq\kappa$, the set $f[\alpha_n]$ has cardinality below $\kappa$. By regularity of $\kappa$, its supremum is below $\kappa$: \begin{align*} \sigma_n=\sup f[\alpha_n]<\kappa. \end{align*} Choose $\alpha_{n+1}<\kappa$ such that \begin{align*} \max\{\alpha_n,\sigma_n\}<\alpha_{n+1}. \end{align*} Now put \begin{align*} \alpha=\sup_{n<\omega}\alpha_n. \end{align*} Because $\omega<\kappa$ and $\kappa$ is regular, this supremum is still below $\kappa$, so $\alpha<\kappa$. Also $\beta<\alpha_0\leq\alpha$. If $\xi<\alpha$, then $\xi<\alpha_n$ for some $n<\omega$, since $\alpha$ is the supremum of the increasing sequence $(\alpha_n)_{n<\omega}$. Therefore \begin{align*} f(\xi)\leq \sup f[\alpha_n]=\sigma_n<\alpha_{n+1}\leq\alpha. \end{align*} Thus $f[\alpha]\subseteq\alpha$, so $\alpha\in C_f$. Since this construction starts above any $\beta<\kappa$, the set $C_f$ is unbounded in $\kappa$. Now we prove closure. Let $\langle \alpha_i:i<\lambda\rangle$ be an increasing sequence from $C_f$, where $0<\lambda<\kappa$, and suppose \begin{align*} \alpha=\sup_{i<\lambda}\alpha_i<\kappa. \end{align*} Take any $\xi<\alpha$. Since the sequence is cofinal in $\alpha$, there is some $i<\lambda$ with $\xi<\alpha_i$. Because $\alpha_i\in C_f$, we have $f[\alpha_i]\subseteq\alpha_i$, hence \begin{align*} f(\xi)<\alpha_i\leq\alpha. \end{align*} Since this holds for every $\xi<\alpha$, we get $f[\alpha]\subseteq\alpha$, so $\alpha\in C_f$. Therefore $C_f$ is closed and unbounded in $\kappa$. This is the basic closure-point mechanism used later to turn a construction into a club set. [/example] The example shows that clubs arise naturally as closure points of functions and constructions. To say that a set is large for reflection purposes, it should meet every such closure set, not merely have cardinality $\kappa$. This motivates the stationary sets, which are the subsets unavoidable by club approximation. [definition: Stationary Set] Let $\kappa$ be a regular uncountable cardinal. A set $S\subset\kappa$ is stationary in $\kappa$ if $S\cap C\neq\varnothing$ for every club $C\subset\kappa$. [/definition] A subset of $\kappa$ can have cardinality $\kappa$ and still fail to be stationary, since it may avoid a club. Stationarity is therefore a structural largeness notion, tuned to transfinite recursion and reflection. The next definition uses it to demand that inaccessible stages appear inside every closed unbounded approximation to $\kappa$. [definition: Mahlo Cardinal] A regular uncountable cardinal $\kappa$ is Mahlo if the set of strongly inaccessible cardinals below $\kappa$ is stationary in $\kappa$. [/definition] This definition directly builds reflection into the hierarchy: a Mahlo cardinal does not merely sit above some inaccessible cardinals, it sees them inside every club approximation to itself. Since every stationary subset of a regular uncountable cardinal is unbounded, Mahloness also gives many inaccessibles below $\kappa$. That abundance should force $\kappa$ itself to inherit the strong limit behaviour of the smaller inaccessible cardinals. [quotetheorem:7401] [citeproof:7401] This result shows why Mahloness is already stronger than inaccessibility: it turns the strong limit property of many smaller cardinals into the strong limit property of $\kappa$ itself. For any $\lambda<\kappa$, stationarity gives a strongly inaccessible $\delta$ with $\lambda<\delta<\kappa$; then $2^\lambda<\delta<\kappa$. Thus the proof is not an abstract appeal to largeness, but a direct conversion of stationary many smaller universes into a cardinal-arithmetic closure property at the top. The stationarity assumption is stronger than mere unboundedness in a way that matters for reflection. Unboundedly many inaccessible cardinals below $\kappa$ would be enough for the strong-limit conclusion, but it could miss a particular club of closure points arising from a construction. Stationarity says this cannot happen, which is why Mahlo cardinals become useful not only as stronger inaccessibles, but as cardinals where arguments can be reflected down to inaccessible closure stages. The theorem does not imply that every inaccessible cardinal below a Mahlo cardinal is itself Mahlo, nor that the inaccessible cardinals form a club. It only guarantees that they are unavoidable by club sets. After establishing this strength comparison, we need the usable form of the Mahlo hypothesis for later reflection arguments: when a construction presents a specific club of closure points, the inaccessible stages meet that club. [quotetheorem:7402] [citeproof:7402] This operational form turns Mahloness into a method. To reflect a construction, first package the closure requirements into a club, then intersect that club with the stationary set of inaccessible cardinals. The hypothesis is exactly tuned to this method: if the inaccessible cardinals were only unbounded, they could still avoid the particular club produced by the construction. The conclusion is also limited in scope; it produces an inaccessible point respecting one specified club, not a single inaccessible point that simultaneously meets an arbitrary proper class of later requirements. The next example makes the method explicit. [example: Reflection Through a Club] Let $\kappa$ be Mahlo, and let $f:\kappa\to\kappa$ code a construction whose closure points are \begin{align*} C_f=\{\alpha<\kappa:f[\alpha]\subset\alpha\}. \end{align*} Assume, as in the closure-point construction, that $C_f$ is club in $\kappa$. By the definition of Mahloness, the set \begin{align*} S=\{\delta<\kappa:\delta\text{ is strongly inaccessible}\} \end{align*} is stationary in $\kappa$. Since $C_f$ is club and $S$ is stationary, we have \begin{align*} S\cap C_f\neq\varnothing. \end{align*} Choose $\delta\in S\cap C_f$. Then $\delta\in S$, so $\delta$ is strongly inaccessible; also $\delta\in C_f$, so \begin{align*} f[\delta]\subset\delta. \end{align*} Because $\delta$ is strongly inaccessible, the usual rank-initial closure argument gives \begin{align*} V_\delta\models\mathrm{ZFC}. \end{align*} The closure condition $f[\delta]\subset\delta$ means that whenever $\xi<\delta$, the stage of the construction coded by $f(\xi)$ is still below $\delta$. Thus the portion of the construction indexed below $\delta$ never leaves the rank-initial universe $V_\delta$. This is the basic reflection pattern: a construction carried out below the Mahlo cardinal $\kappa$ has a strongly inaccessible closure point $\delta<\kappa$, so it reflects down to the smaller universe $V_\delta$. [/example] ## Consistency Strength and Relative Consistency The final problem in this chapter is methodological. Large cardinal axioms usually cannot be proved from ZFC, and their outright consistency cannot be proved in ZFC if ZFC is consistent. Set theorists therefore compare theories by relative consistency rather than by absolute proof from the base theory. [definition: Relative Consistency] Let $T$ and $S$ be formal theories. The theory $T$ is relatively consistent with $S$ if $\operatorname{Con}(S)$ implies $\operatorname{Con}(T)$. [/definition] This definition records a metamathematical implication, not a theorem of the stronger theory alone. In practice, relative consistency is usually established by constructing a model of $T$ from a model of $S$. The first question is what extra consistency follows from assuming an inaccessible cardinal. The inaccessible rank theorem answers this by producing a set-sized universe of ZFC inside the larger model. [quotetheorem:7403] [citeproof:7403] The distinction between these two conclusions matters. A non-well-founded model can believe that it has a rank-initial model of ZFC without giving the metatheory an externally transitive model. Thus the theorem does not say that ordinary consistency of ZFC plus an inaccessible cardinal automatically yields a transitive model of ZFC; that stronger conclusion requires a well-foundedness hypothesis. What it does give is the standard relative-consistency comparison that survives Gödelian limitations, and this is enough to place inaccessible cardinals above bare ZFC in consistency strength. [remark: Consistency Strength as a Hierarchy] When large cardinal axiom $A$ proves the consistency of large cardinal axiom $B$ over a common base theory, and no reverse implication is known or expected, $A$ is regarded as having greater consistency strength than $B$. The hierarchy is not merely a list of stronger-looking definitions; it is organised by the models and inner models that each axiom can produce. [/remark] The first levels now fit together as follows: weakly inaccessible cardinals express regular limit size, strongly inaccessible cardinals make $V_\kappa$ into a model of ZFC, and Mahlo cardinals require stationarily many inaccessible stages below them. Chapter 2 replaces stationary reflection by logical reflection and indescribability; Chapters 3 and 4 then replace these reflection principles by measures, ultrapowers, and elementary embeddings. The first chapter introduced the early hierarchy through closure, inaccessibility, and stationary reflection. We now sharpen that perspective by asking not just whether many smaller stages are large, but which logical truths about $V_\kappa$ must already appear below $\kappa$. # 2. Reflection and Indescribability Reflection is the mechanism by which the universe of sets reveals finite pieces of its structure inside smaller ranks. In the first chapter, inaccessible cardinals entered as ranks $V_\kappa$ that satisfy large fragments, and sometimes all, of ZFC. This chapter asks a sharper question: which statements about $V$ or about $V_\kappa$ must already be true in some smaller $V_\alpha$? The answer begins with first-order reflection in ZFC and leads to second-order reflection principles strong enough to characterize weak compactness. ## The Levy Hierarchy and Formulas over $V_\kappa$ To discuss reflection, we need a way to measure the logical complexity of a statement. The important distinction is not its length but the pattern of unbounded quantifiers over the universe, since bounded quantifiers are already local to a set. [definition: Bounded Quantifier] A quantifier is bounded if it has the form $\forall x \in y$ or $\exists x \in y$, and a formula is $\Delta_0$ if every quantifier occurring in it is bounded. [/definition] Bounded formulas are the formulas whose truth is controlled by the sets already named in the formula. They form the base level of the Levy hierarchy. The next task is to sort formulas by how many unbounded quantifiers must be added above this base. [definition: Levy Hierarchy] For $n \in \mathbb N$, the classes $\Sigma_n$, $\Pi_n$, and $\Delta_n$ of first-order formulas in the language of set theory are defined recursively as follows: - $\Sigma_0 = \Pi_0 = \Delta_0$ is the class of bounded formulas. - $\Sigma_{n+1}$ is the smallest class of formulas containing all formulas $\exists x\,\varphi(x)$ with $\varphi \in \Pi_n$ and closed under finite conjunctions, finite disjunctions, and bounded quantification. - $\Pi_{n+1}$ is the smallest class of formulas containing all formulas $\forall x\,\varphi(x)$ with $\varphi \in \Sigma_n$ and closed under finite conjunctions, finite disjunctions, and bounded quantification. - $\Delta_{n+1}$ is the class of formulas belonging to both $\Sigma_{n+1}$ and $\Pi_{n+1}$ over the relevant background theory. [/definition] In practice, a formula may be rewritten into an equivalent prenex or block normal form before its level is read off. The closure clauses in the definition are what make the hierarchy stable under the bounded quantifiers and finite Boolean combinations that occur throughout ordinary set-theoretic definitions. The hierarchy is useful because many set-theoretic properties have a stable complexity. For example, existence assertions tend to be $\Sigma_1$, while universal closure properties tend to be $\Pi_1$ or higher. Before we reflect formulas to smaller ranks, it helps to see how a low-complexity statement separates the witness from the bounded verification. [example: A Sigma One Assertion] Let $R$ be a definable binary relation on a set $a$. The assertion “there is a rank-initial transitive set containing $a$ in which every $R$-chain has an upper bound” has the form $\exists x\,\psi(x,a,R)$: the variable $x$ is the proposed rank-initial transitive set, and $\psi$ says that $a \in x$, that $x$ has the chosen rank-initial transitive-set code, and that the following bounded closure condition holds inside $x$. For a set $c$, write “$c$ is an $R$-chain in $a$” as the bounded condition \begin{align*} c \subseteq a \ \wedge\ \forall u \in c\,\forall v \in c\,\bigl(u=v \vee (u,v)\in R \vee (v,u)\in R\bigr). \end{align*} The upper-bound clause for such a chain is \begin{align*} \exists b \in x\,\bigl(b \in a \wedge \forall u \in c\, (u,b)\in R\bigr). \end{align*} Thus the closure part of $\psi(x,a,R)$ can be written as \begin{align*} \forall c \in x\,\Bigl(c \subseteq a \wedge \forall u \in c\,\forall v \in c\,\bigl(u=v \vee (u,v)\in R \vee (v,u)\in R\bigr) \rightarrow \exists b \in x\,\bigl(b \in a \wedge \forall u \in c\, (u,b)\in R\bigr)\Bigr). \end{align*} Once $x$, $a$, and $R$ are named, every quantifier in this verification is bounded by $x$, $a$, or $c$. The only unbounded quantifier in the whole assertion is the initial existential quantifier $\exists x$, so the statement is $\Sigma_1$ in the Levy hierarchy. A rank $V_\alpha$ may contain the parameters $a$ and $R$ while failing to contain any such witness $x$; the point of $\Sigma_1$ reflection is precisely that existential witnesses can live above the rank that already contains the parameters. [/example] This example shows why rank initial segments are the natural structures for reflection: they contain parameters and may or may not contain the witnesses demanded by unbounded quantifiers. To state reflection precisely, we must specify how a formula originally interpreted in $V$ is interpreted inside $V_\kappa$. That leads to relativization. [definition: Relativization to $V_\kappa$] Let $\kappa$ be an ordinal. Relativization to $V_\kappa$ is the syntactic map sending each first-order formula $\varphi$ in the language of set theory to a formula $\varphi^{V_\kappa}$ in which every unbounded quantifier is bounded by $V_\kappa$: each $\exists x$ is replaced by $\exists x \in V_\kappa$, and each $\forall x$ is replaced by $\forall x \in V_\kappa$. [/definition] Relativization turns a statement about the universe into a statement about the structure $(V_\kappa,\in)$. When $a_1,\dots,a_m \in V_\kappa$, we write \begin{align*} V_\kappa \models \varphi(a_1,\dots,a_m) \end{align*} for the truth of the relativized formula in this structure. The role of parameters must be tracked because a formula cannot reflect below the rank of the sets it names. [remark: Parameters Matter] Reflection statements almost always include parameters. If $a \in V_\kappa$, a reflected instance of $\varphi(a)$ must be found inside some $V_\alpha$ with $a \in V_\alpha$, so the reflecting rank must lie above the ranks of the parameters. [/remark] ## Reflection Principles in ZFC The guiding problem is whether the universe can be distinguished from all of its initial segments by a single first-order sentence. ZFC proves that no finite list of formulas can do this: each finite fragment of first-order truth in $V$ reflects down to arbitrarily large ranks. [quotetheorem:4844] [citeproof:4844] This theorem is striking because it is a theorem of ZFC rather than a large cardinal assumption. It is also deliberately finite. The finite hypothesis cannot be replaced by a single set-sized rank reflecting the whole first-order language: by the Levy-Montague obstruction, there is no ordinal $\alpha$ such that $(V_\alpha,\in) \prec (V,\in)$ for all first-order formulas with parameters from $V_\alpha$. Equivalently, the would-be strengthening \begin{align*} \exists \alpha\ \forall \varphi\ \forall a_1,\dots,a_m \in V_\alpha\, \bigl(V \models \varphi(a_1,\dots,a_m) \iff V_\alpha \models \varphi(a_1,\dots,a_m)\bigr) \end{align*} is not a theorem of ZFC; in fact it is impossible. The obstruction is a uniformity obstruction, not a failure of the finite theorem. For each finite list $\Phi$ we may build the finitely many Skolem functions needed for $\Phi$, but the collection of all first-order formulas would require a set-sized rank to carry a uniform truth-agreement scheme for set theory. Tarski's undefinability theorem is the diagonal reason behind this limitation: sufficiently rich set-theoretic truth cannot be captured by one internally available truth predicate. Levy reflection therefore supplies arbitrarily large mirrors for any finite fragment chosen in advance, not one rank that mirrors the whole universe at once. The parameter condition has its own sharp obstruction. If $a \notin V_\alpha$, then the formula $x=a$ cannot even be interpreted in $V_\alpha$ with $a$ as a parameter; if $a \in V_\alpha$ but a statement asserts $\exists y(a \in y \wedge y \text{ is transitive})$, the reflected witness must also lie in the reflecting rank. Levy reflection handles such parameters by demanding agreement only for parameters already in $V_\alpha$ and by closing under the relevant Skolem functions for all tuples from $V_\alpha$. Thus the theorem gives arbitrarily large mirrors for each finite fragment chosen in advance, while the full reflection scheme remains a proper class of separate assertions rather than one set-sized elementary rank. The following example turns this finite-fragment reflection into a concrete closure assertion. [example: Reflecting a First-Order Closure Statement] Fix $n \in \mathbb N$, and let $F:V^n \to V$ be defined by the first-order formula $\theta(z,u_1,\dots,u_n)$, so that $V$ satisfies \begin{align*}\forall u_1\cdots\forall u_n\,\exists!z\,\theta(z,u_1,\dots,u_n).\end{align*} The closure assertion can be written as the single first-order sentence \begin{align*}\forall x\,\exists y\,\bigl(x\subseteq y \wedge y\text{ is transitive}\wedge \forall u_1\in y\cdots\forall u_n\in y\,\exists z\in y\,\theta(z,u_1,\dots,u_n)\bigr).\end{align*} Choose any ordinal $\gamma$. By the Levy reflection theorem quoted above, applied to this displayed formula together with the defining formula $\theta$, there is an ordinal $\alpha>\gamma$ such that truth of these formulas agrees between $V$ and $V_\alpha$ for parameters from $V_\alpha$. Therefore, if $x\in V_\alpha$, then $V\models \exists y\,\chi(x,y)$, where \begin{align*}\chi(x,y)\equiv x\subseteq y \wedge y\text{ is transitive}\wedge \forall u_1\in y\cdots\forall u_n\in y\,\exists z\in y\,\theta(z,u_1,\dots,u_n).\end{align*} Reflection gives \begin{align*}V_\alpha\models \exists y\in V_\alpha\,\chi(x,y).\end{align*} Thus the witness $y$ is not merely a set existing somewhere in $V$; it may be chosen with $y\in V_\alpha$, and inside $V_\alpha$ it contains $x$, is transitive, and is closed under the operation defined by $\theta$. The example shows how a fixed first-order closure property reflects to arbitrarily large ranks once the arity and defining formula of the class function have been fixed. [/example] The example reflects from $V$ downward to some rank chosen after the formula is fixed. Large cardinal reflection asks for a fixed cardinal $\kappa$ that reflects statements from $V_\kappa$ to smaller ranks below itself. We therefore isolate the scheme of first-order reflection at a cardinal. [definition: First-Order Reflection at a Cardinal] Let $\kappa$ be an uncountable cardinal. We say that $\kappa$ has first-order reflection for a class $\Gamma$ of formulas if for every $\varphi(x_1,\dots,x_m) \in \Gamma$ and all parameters $a_1,\dots,a_m \in V_\kappa$, whenever \begin{align*} V_\kappa \models \varphi(a_1,\dots,a_m), \end{align*} there is some $\alpha < \kappa$ with $a_1,\dots,a_m \in V_\alpha$ such that \begin{align*} V_\alpha \models \varphi(a_1,\dots,a_m). \end{align*} [/definition] This definition fixes the ambient universe as $V_\kappa$ rather than $V$. Reflection at a cardinal becomes stronger when the target $\alpha$ must lie below the same $\kappa$ for every relevant statement. The most useful ranks for this purpose are inaccessible ranks, since they support enough set-theoretic closure to resemble universes. [remark: Why Inaccessibility Appears] If $\kappa$ is inaccessible, then $V_\kappa$ has enough closure to behave like a small universe of set theory. This makes reflection from $V_\kappa$ to $V_\alpha$ meaningful: both sides are rank-initial structures, and the lower rank can inherit substantial fragments of the higher rank's theory. [/remark] ## $\Pi^1_n$ Indescribability and Weak Compactness First-order reflection does not see arbitrary subsets of $V_\kappa$. Large cardinal strength appears when we allow second-order parameters, such as subsets $A \subset V_\kappa$, and ask whether properties of the expanded structure $(V_\kappa,\in,A)$ reflect to some smaller $(V_\alpha,\in,A \cap V_\alpha)$. [definition: Second-Order Formula over $V_\kappa$] A second-order formula over $V_\kappa$ is a formula interpreted in the structure $(V_\kappa,\in)$ that may quantify over subsets of $V_\kappa$ as second-order variables. [/definition] Second-order quantifiers increase the strength of reflection because they can encode combinatorial objects on $\kappa$, such as trees, colorings, and clubs. To compare different second-order reflection principles, we need a hierarchy counting alternations of second-order quantifiers. This is the projective hierarchy over $V_\kappa$. [definition: Projective Hierarchy over $V_\kappa$] A formula is $\Pi^1_n$ over $V_\kappa$ if it begins with $n$ alternating blocks of second-order quantifiers, the first block universal, followed by a first-order formula over $(V_\kappa,\in)$; a formula is $\Sigma^1_n$ over $V_\kappa$ if it begins with $n$ alternating blocks of second-order quantifiers, the first block existential, followed by a first-order formula over $(V_\kappa,\in)$. [/definition] This hierarchy identifies the formulas whose reflection strength we want to measure. The universal first block of $\Pi^1_n$ is the setting for saying that a rank cannot be described by a property that all smaller ranks fail. This motivates the next definition, which names cardinals reflecting $\Pi^1_n$ sentences with arbitrary predicate parameters. [definition: $\Pi^1_n$ Indescribable Cardinal] Let $n \in \mathbb N$ and let $\kappa$ be an inaccessible cardinal. The cardinal $\kappa$ is $\Pi^1_n$ indescribable if for every $A \subset V_\kappa$ and every $\Pi^1_n$ sentence $\varphi$ in the language with an additional predicate symbol, \begin{align*} (V_\kappa,\in,A) \models \varphi \end{align*} implies that there exists an ordinal $\alpha < \kappa$ such that \begin{align*} (V_\alpha,\in,A \cap V_\alpha) \models \varphi. \end{align*} [/definition] The reflecting ordinal $\alpha$ is not built into the definition as inaccessible. Some courses and texts study stronger variants where the reflecting targets are required to have additional closure properties, but the standard indescribability notion only asks for some smaller rank. The point is that even a property allowed to quantify over all subsets of $V_\kappa$ cannot uniquely describe $V_\kappa$ among the lower ranks. For instance, a predicate $A \subset V_\kappa$ can code a $\kappa$-tree or a coloring of $[\kappa]^2$, objects invisible to ordinary first-order reflection with set parameters alone. At level $\Pi^1_1$, this second-order reflection is tied to a compactness property for trees. We now introduce the weak compactness formulation that the course uses for this connection. [definition: Weakly Compact Cardinal] An uncountable cardinal $\kappa$ is weakly compact if it is inaccessible and satisfies the tree property: every $\kappa$-tree has a cofinal branch. [/definition] This definition uses the tree property as the course's combinatorial entry point. Other equivalent definitions use compactness of infinitary logic or partition relations. The next theorem explains why weak compactness is also a reflection principle: tree compactness prevents failures of $\Pi^1_1$ reflection from cohering through all levels below $\kappa$. [quotetheorem:7404] [proofunderconstruction:7404] This characterization shows the general pattern behind indescribability arguments: nonreflection produces a coherent system of local counterexamples, and weak compactness turns that system into a global counterexample. The converse shows why this is a characterization rather than only an implication: a branch through a tree is itself a second-order object, so the assertion that no such branch exists is exactly the kind of universal second-order statement that $\Pi^1_1$ indescribability reflects. The hypotheses are doing real work. Without inaccessibility, the coding and size requirements can fail in concrete ways. At $\omega_1$, which is regular but not inaccessible, ZFC proves that there is an $\omega_1$-Aronszajn tree; this is a tree of height $\omega_1$ with countable levels and no uncountable branch, so the tree-property conclusion fails outright. At a regular successor $\kappa=\lambda^+$, the full binary tree $2^{<\kappa}$ illustrates the size obstruction: if $2^\lambda \ge \kappa$, then some bounded level already has size at least $\kappa$, so the usual coding tree is no longer a $\kappa$-tree and the tree property, even if considered separately, would not apply to it. Without the tree property, the local counterexamples may form an Aronszajn tree: every bounded approximation exists, but no cofinal branch assembles them into one global object. The theorem also does not say that every second-order statement reflects; it is restricted to the $\Pi^1_1$ pattern, and higher alternations of second-order quantifiers lead to stronger indescribability notions. The next example illustrates what is reflected: not just sets, but universal assertions about all subsets of a rank. [example: A Reflected Universal Second-Order Assertion] Let $A \subset V_\kappa$ code the extra structure, and write the universal second-order assertion in the explicit $\Pi^1_1$ form \begin{align*} (V_\kappa,\in,A)\models \forall X\,\bigl(\chi(X,A)\rightarrow \exists \beta\,\rho(X\cap V_\beta,A\cap V_\beta,\beta)\bigr), \end{align*} where $\chi(X,A)$ is the first-order closure property and $\rho(X\cap V_\beta,A\cap V_\beta,\beta)$ says that $X\cap V_\beta$ is the required compatible rank-initial approximation. Here the quantifier $\forall X$ ranges over all subsets $X\subset V_\kappa$, while the formulas $\chi$ and $\rho$ use only first-order quantifiers over the ambient rank. Since $\kappa$ is weakly compact, the weak-compactness characterization above gives $\Pi^1_1$ indescribability of $\kappa$. Applying it to the predicate $A$ and the displayed $\Pi^1_1$ sentence yields some ordinal $\alpha<\kappa$ such that \begin{align*} (V_\alpha,\in,A\cap V_\alpha)\models \forall X\,\bigl(\chi(X,A\cap V_\alpha)\rightarrow \exists \beta\,\rho(X\cap V_\beta,A\cap V_\beta,\beta)\bigr). \end{align*} In the reflected structure, the same displayed quantifier $\forall X$ now ranges over all subsets $X\subset V_\alpha$, and any witness $\beta$ is interpreted inside $V_\alpha$, so necessarily $\beta<\alpha$. Thus the reflected statement is not just about the particular predicate $A\cap V_\alpha$; it is a universal assertion about every subset of the smaller rank. The ordinal $\alpha$ need not be inaccessible, because $\Pi^1_1$ indescribability asks only for some reflecting $\alpha<\kappa$. [/example] ## Trees, Branches, and Partition Properties The next problem is to translate reflection into concrete combinatorics on $\kappa$. Trees are the central test case: a tree property says that a coherent system of approximations of height $\kappa$ cannot avoid producing a full object of length $\kappa$. Visually, one should imagine narrow levels indexed by ordinals below $\kappa$, with bounded branches representing compatible partial choices; the tree property says that at a weakly compact cardinal these bounded approximations assemble into a branch whose heights are cofinal in $\kappa$. [definition: $\kappa$-Tree] Let $\kappa$ be an uncountable cardinal. A $\kappa$-tree is a partially ordered set $(T,<_{T})$ such that: - for each $t \in T$, the set $\{s \in T : s <_T t\}$ is well-ordered by $<_T$; - $T$ has height $\kappa$; - each level $T_\alpha = \{t \in T : \operatorname{height}_T(t)=\alpha\}$ has cardinality less than $\kappa$. [/definition] A cofinal branch is a linearly ordered subset meeting levels cofinally in $\kappa$. A $\kappa$-Aronszajn tree is a $\kappa$-tree with no cofinal branch. These two notions identify the obstruction that the tree property forbids. [definition: Tree Property] An uncountable cardinal $\kappa$ has the tree property if there is no $\kappa$-Aronszajn tree. [/definition] At $\omega_1$, Aronszajn trees exist in ZFC, so the tree property cannot hold at every regular uncountable cardinal. The large-cardinal phenomenon is that inaccessible cardinals with the tree property sit at the first weakly compact level. The contrast with $\omega_1$ is the standard example. [example: Aronszajn Trees versus the Tree Property] An $\omega_1$-Aronszajn tree is a tree $T$ of height $\omega_1$ such that each level $T_\alpha$ is countable and no branch meets cofinally many levels. Since $\omega_1$ is regular, a subset of $\omega_1$ is cofinal in $\omega_1$ exactly when it is uncountable. Therefore every branch through an $\omega_1$-Aronszajn tree is bounded in height, hence has countably many levels and is countable. This shows that regularity of the height cardinal does not by itself force a cofinal branch through a tree with small levels: the levels of $T$ are all smaller than $\omega_1$, but there is still no uncountable branch. By contrast, if $\kappa$ is weakly compact, then by definition $\kappa$ is inaccessible and has the tree property, so every $\kappa$-tree has a cofinal branch. Thus the Aronszajn obstruction that occurs at $\omega_1$ cannot occur at a weakly compact cardinal. [/example] The example identifies tree compactness as a branching principle. To compare it with Ramsey theory, we need a way to state the analogous obstruction for colorings: a coloring of pairs may try to split every large set into two colors, and compactness should prevent that from happening. The partition notation below names the demand that every two-coloring of pairs from $\kappa$ has a homogeneous subset still as large as $\kappa$. [definition: Partition Relation $\kappa \to (\kappa)^2_2$] For a cardinal $\kappa$, the relation $\kappa \to (\kappa)^2_2$ means that for every coloring $c:[\kappa]^2 \to \{0,1\}$, there is a set $H \subset \kappa$ of cardinality $\kappa$ such that $c$ is constant on $[H]^2$. [/definition] The partition relation is stronger than the finite Ramsey theorem in both size and conclusion: it asks for a homogeneous set as large as the original cardinal. The theorem below records the equivalence between the tree formulation and the usual weak compactness formulations when $\kappa$ is inaccessible. [quotetheorem:7405] [citeproof:7405] Because this page defined weak compactness as inaccessibility plus the tree property, the tree formulation is the starting point rather than a new consequence. The theorem is still useful because it explains why the same cardinal property appears in three languages: trees, Ramsey colorings, and infinitary compactness. Each formulation also has a boundary. The tree property concerns only trees with levels of size less than $\kappa$; it says nothing about wider trees, and it does not rule out Aronszajn trees at unrelated cardinals. The partition relation here is only the pair relation $\kappa \to (\kappa)^2_2$; it is not the full Ramsey-cardinal conclusion for all finite arities. The compactness clause is $\kappa$-compactness for $L_{\kappa,\kappa}$ in languages and theories of size at most $\kappa$, not unrestricted compactness for arbitrary-sized theories or stronger logics. The inaccessibility hypothesis cannot be dropped from these translations: at successors, level-size estimates may fail, and at $\omega_1$ there are Aronszajn trees in ZFC, so regularity alone does not supply the compactness phenomenon. The partition formulation will be used next to extract homogeneous subsets from colorings, so we record the direction from weak compactness to pairs as its own theorem. [quotetheorem:7406] [citeproof:7406] Partition properties make weak compactness feel close to Ramsey cardinals, but the strength is lower: only pairs and two colors are required, not arbitrary finite subsets at all finite arities. The hypotheses again matter. Sierpiński's theorem gives a coloring \begin{align*} c:[\omega_1]^2 \to \{0,1\} \end{align*} with no uncountable homogeneous set, usually written \begin{align*} \omega_1 \nrightarrow (\omega_1)^2_2. \end{align*} This is the precise counterexample to replacing weak compactness by regular uncountability: $\omega_1$ is regular, but it fails the pair partition conclusion. It also matches the tree boundary, since $\omega_1$-Aronszajn trees exist in ZFC. Inaccessibility is what lets the bounded approximation tree stay narrow; without it, a construction may have levels of size $\kappa$ and the tree property gives no contradiction. The pair partition property also does not by itself make $\kappa$ a Ramsey cardinal, since Ramsey cardinals require homogeneous sets for colorings of all finite subsets and allow stronger measures of coherence. The final example shows how the branch produced by the tree property becomes the homogeneous set promised by the partition relation. [example: From a Branch to a Homogeneous Set] Suppose the tree construction for $c:[\kappa]^2\to\{0,1\}$ produces a cofinal branch $B$ whose nodes are homogeneous approximations of color $0$, ordered by end-extension. For each node $s\in B$, let $A_s\subset\kappa$ be the set of ordinals recorded by $s$, and assume that the construction records homogeneity by requiring \begin{align*} \forall \alpha\in A_s\,\forall \beta\in A_s\,(\alpha\neq\beta\rightarrow c(\{\alpha,\beta\})=0). \end{align*} Define \begin{align*} H=\bigcup_{s\in B}A_s. \end{align*} Because $B$ is cofinal in the tree of height $\kappa$, for every $\xi<\kappa$ there is a node $s\in B$ whose level is above $\xi$. The approximation at that level contains ordinals arbitrarily high below $\kappa$, so $H$ is unbounded in $\kappa$. Since $\kappa$ is inaccessible, in particular regular, every unbounded subset of $\kappa$ has cardinality $\kappa$; hence $|H|=\kappa$. It remains to verify homogeneity. Let $\alpha,\beta\in H$ with $\alpha\neq\beta$. By the definition of $H$, choose nodes $s,t\in B$ such that $\alpha\in A_s$ and $\beta\in A_t$. Since $B$ is a branch, its nodes are linearly ordered by end-extension, so either $s\leq t$ or $t\leq s$. If $s\leq t$, then end-extension gives $A_s\subseteq A_t$, hence $\alpha,\beta\in A_t$, and the homogeneity condition for $t$ gives \begin{align*} c(\{\alpha,\beta\})=0. \end{align*} If $t\leq s$, then $A_t\subseteq A_s$, hence $\alpha,\beta\in A_s$, and the homogeneity condition for $s$ gives the same conclusion. Thus every pair from $H$ has color $0$, so the cofinal branch decodes exactly into a homogeneous subset of $\kappa$ of size $\kappa$. [/example] The chapter's progression can now be summarized as a ladder of reflection strength. ZFC gives first-order reflection from $V$ to arbitrarily large $V_\alpha$. Inaccessible cardinals, introduced in Chapter 1, make individual ranks $V_\kappa$ into set-sized universes. Weakly compact cardinals add second-order reflection, tree compactness, and partition regularity, forming the first major reflection principle beyond Mahloness. The next chapter changes the language again: instead of asking which statements reflect downward, it asks when a cardinal carries a complete ultrafilter that measures all subsets of it. Indescribability turns reflection into a precise logical and combinatorial strength. The next chapter changes the organizing device from formulas and stationary sets to ultrafilters, where largeness is encoded by a complete measure on all subsets of a cardinal. # 3. Ultrafilters and Measurable Cardinals This chapter moves from the combinatorial strength of inaccessible and Mahlo cardinals to the measure-theoretic formulation of largeness. The guiding question is how a cardinal can carry a $0$-$1$ notion of largeness that is closed under long intersections. Ultrafilters on $\kappa$ turn subsets of $\kappa$ into large or small sets, and measurable cardinals are exactly those cardinals where such a measure exists in a highly complete, nontrivial form. The chapter also introduces normality, the closure condition that makes ultrafilters interact well with ordinal-indexed diagonal arguments. Normal measures convert local choices below $\kappa$ into a single coherent measure-one set, and this is the source of the first major consequences: measurables are inaccessible, Mahlo, and support a version of Fodor's lemma internal to the measure. ## Complete Ultrafilters and Nonprincipal Measures The starting problem is to formalise a notion of "almost all ordinals below $\kappa$" that is strong enough to survive intersections indexed by all smaller cardinals. Ordinary countable completeness is too weak for large cardinal arguments at $\kappa$, because the natural constructions make fewer than $\kappa$ many choices at once. This leads to $\kappa$-complete filters. [definition: Kappa Complete Filter] Let $\kappa$ be an uncountable cardinal. A filter $U$ on $\kappa$ is $\kappa$-complete if for every $\lambda < \kappa$ and every sequence $(A_i)_{i<\lambda}$ of elements of $U$, we have \begin{align*} \bigcap_{i<\lambda} A_i \in U. \end{align*} [/definition] A complete filter can say that many sets are large, but it need not decide every subset of $\kappa$. For a two-valued measure, every set should be either large or have large complement, so the next requirement is maximality as a filter. [definition: Kappa Complete Ultrafilter] Let $\kappa$ be an uncountable cardinal. A $\kappa$-complete ultrafilter on $\kappa$ is an ultrafilter $U$ on $\kappa$ that is $\kappa$-complete. [/definition] Maximality alone permits a degenerate example: the ultrafilter of all sets containing a fixed ordinal. Such an ultrafilter decides every subset of $\kappa$, but it does so by asking only whether one selected point belongs to the set, so it carries no global information about $\kappa$. The definition below rules out this point-mass behaviour by requiring every singleton to be negligible. [definition: Nonprincipal Measure] Let $\kappa$ be an uncountable cardinal. A nonprincipal measure on $\kappa$ is a $\kappa$-complete ultrafilter $U$ on $\kappa$ such that no singleton belongs to $U$. [/definition] Nonprincipality makes every co-singleton large, and completeness then makes small unions of points negligible. The first structural test for the definition is therefore whether the measure discards all bounded subsets of $\kappa$. [quotetheorem:7407] [citeproof:7407] The theorem uses all three hypotheses in visible ways. Nonprincipality rules out singletons; without it, a principal ultrafilter would declare one singleton large. $\kappa$-completeness is what lets us pass from singletons to arbitrary sets of size below $\kappa$; mere finite completeness would only discard finite sets. The conclusion is also deliberately limited: it says nothing about subsets of size $\kappa$, which are exactly the sets the ultrafilter is meant to decide in a nontrivial way. [example: Cofinal Tail Sets] Let $\kappa$ carry a nonprincipal $\kappa$-complete ultrafilter $U$, and for $\alpha<\kappa$ put $T_\alpha=\{\beta<\kappa:\beta>\alpha\}$. Since $T_\alpha=\kappa\setminus(\alpha+1)$ and $\alpha+1$ has cardinality less than $\kappa$, the bounded-set conclusion of the bounded-set theorem above gives $T_\alpha\in U$. For a finite list $\alpha_0,\ldots,\alpha_{n-1}<\kappa$, let $m=\max\{\alpha_i:i<n\}$. Then for every $\beta<\kappa$, \begin{align*} \beta\in T_{\alpha_0}\cap\cdots\cap T_{\alpha_{n-1}} \text{ exactly when } \beta>\alpha_i \text{ for every } i<n. \end{align*} Because $m$ is one of the $\alpha_i$ and every $\alpha_i\le m$, this condition is equivalent to $\beta>m$. Hence \begin{align*} T_{\alpha_0}\cap\cdots\cap T_{\alpha_{n-1}}=T_m. \end{align*} In particular this finite intersection belongs to $U$. More generally, let $(\alpha_i)_{i<\lambda}$ have length $\lambda<\kappa$. Each $T_{\alpha_i}$ belongs to $U$, so $\kappa$-completeness gives \begin{align*} \bigcap_{i<\lambda}T_{\alpha_i}\in U. \end{align*} If $\delta=\sup_{i<\lambda}\alpha_i$ is still below $\kappa$, then every $\beta\in T_\delta$ satisfies $\beta>\delta\ge\alpha_i$ for all $i<\lambda$, so \begin{align*} T_\delta\subseteq\bigcap_{i<\lambda}T_{\alpha_i}. \end{align*} Thus small intersections of cofinal tails remain measure-one, and when $\kappa$ is known to be regular the supremum $\delta$ is below $\kappa$ for every $\lambda<\kappa$ sequence. The restriction to fewer than $\kappa$ many intersections is essential. If $\beta<\kappa$, then taking $\alpha=\beta$ gives $\beta\notin T_\alpha$, since $\beta>\beta$ is false. Therefore no $\beta<\kappa$ lies in every $T_\alpha$, and \begin{align*} \bigcap_{\alpha<\kappa}T_\alpha=\varnothing. \end{align*} So $\kappa$-completeness preserves all smaller intersections of measure-one tails, but it does not imply closure under $\kappa$ many intersections. [/example] The example shows ordinary intersections preserving tail largeness. The next problem is harder: course arguments often produce $\kappa$ many measure-one requirements, so we need a closure operation weaker than full intersection but still strong enough for transfinite recursion. ## Normal Measures and Diagonal Intersections Many arguments below a large cardinal produce a sequence $(A_\alpha)_{\alpha<\kappa}$ of measure-one sets, one set for each stage. Ordinary $\kappa$-completeness cannot take their full intersection, since the sequence has length $\kappa$. The diagonal intersection keeps the $\alpha$th requirement only after stage $\alpha$, and this is the right closure operation for normal measures. [definition: Diagonal Intersection] Let $\kappa$ be an uncountable cardinal. The diagonal-intersection operation is the map \begin{align*} \Delta_\kappa:(\mathcal P(\kappa))^\kappa\to\mathcal P(\kappa) \end{align*} defined as follows: for a sequence $(A_\alpha)_{\alpha<\kappa}$ of subsets of $\kappa$, \begin{align*} \Delta_\kappa((A_\alpha)_{\alpha<\kappa})=\Delta_{\alpha<\kappa} A_\alpha = \{\xi < \kappa : \xi \in A_\alpha \text{ for all } \alpha < \xi\}. \end{align*} [/definition] The set $\Delta_{\alpha<\kappa} A_\alpha$ asks the ordinal $\xi$ to satisfy all earlier requirements. This is weaker than lying in every $A_\alpha$, but it is strong enough for inductive constructions, because requirements introduced after stage $\xi$ should not constrain $\xi$. [definition: Normal Measure] Let $\kappa$ be an uncountable cardinal. A nonprincipal $\kappa$-complete ultrafilter $U$ on $\kappa$ is normal if for every sequence $(A_\alpha)_{\alpha<\kappa}$ of elements of $U$, the diagonal intersection $\Delta_{\alpha<\kappa} A_\alpha$ belongs to $U$. [/definition] Diagonal intersections are useful for constructing ordinals satisfying earlier constraints, but many applications present the same phenomenon as a statement about functions. The next definition names the functions that always point downward, because these are exactly the functions that normality forces to stabilise on a measure-one set. [definition: Regressive Function] Let $S \subset \kappa$. A function $f:S \to \kappa$ is regressive if $f(\xi) < \xi$ for every nonzero $\xi \in S$. [/definition] A regressive function assigns to each ordinal a smaller witness, so globally it may have many values while locally it is forced below the point being measured. Normality should prevent such a function from varying on a measure-one set forever, and the next theorem makes that intuition exact. [quotetheorem:7408] [citeproof:7408] This equivalence is the operational heart of normality: diagonal closure and constancy of regressive functions are the same phenomenon. The nonprincipality assumption keeps the statement from being vacuous, since a principal ultrafilter would make every function constant on the distinguished singleton. Completeness cannot be treated as a cosmetic hypothesis: if $U$ is the cofinite ultrafilter on $\omega$, then the regressive-function formulation is not a meaningful large-cardinal normality principle, because the partition of $\omega$ into singletons has all fibres $U$-null and ordinary finite completeness gives no way to intersect countably many complements. At $\kappa$, the proof needs the same kind of closure for fewer than $\kappa$ many small exceptional sets, while normality supplies the extra diagonal step that $\kappa$-completeness alone cannot provide. In practice, this theorem lets us replace long diagonal-intersection arguments by the simpler instruction: build a regressive obstruction and force it to be constant on a measure-one set. [example: Diagonal Intersection Avoids a False Full Intersection] Let $U$ be a normal measure on $\kappa$, and define $A_\alpha=\kappa\setminus(\alpha+1)$ for each $\alpha<\kappa$. Since $\alpha+1$ has cardinality less than $\kappa$, the bounded-set theorem above gives \begin{align*} A_\alpha=\kappa\setminus(\alpha+1)\in U. \end{align*} The full intersection has no elements. Indeed, if $\xi<\kappa$, then the requirement for the $\alpha=\xi$ coordinate would be \begin{align*} \xi\in A_\xi=\kappa\setminus(\xi+1). \end{align*} But $\xi\in\xi+1$, so $\xi\notin\kappa\setminus(\xi+1)$. Hence no $\xi<\kappa$ belongs to every $A_\alpha$, and therefore \begin{align*} \bigcap_{\alpha<\kappa}A_\alpha=\varnothing. \end{align*} The diagonal intersection asks for less. By definition, \begin{align*} \Delta_{\alpha<\kappa}A_\alpha=\{\xi<\kappa:\xi\in A_\alpha\text{ for every }\alpha<\xi\}. \end{align*} If $\xi=0$, there is no $\alpha<0$, so the condition is vacuous. If $\xi>0$ and $\alpha<\xi$, then $\xi\notin\alpha+1$ because every element of $\alpha+1$ is at most $\alpha$, while $\xi>\alpha$. Thus \begin{align*} \xi\in\kappa\setminus(\alpha+1)=A_\alpha \end{align*} for every $\alpha<\xi$. Therefore every $\xi<\kappa$ lies in the diagonal intersection, so \begin{align*} \Delta_{\alpha<\kappa}A_\alpha=\kappa. \end{align*} Since $\kappa\in U$, the diagonal intersection is measure-one even though the full $\kappa$-length intersection is empty; the diagonal version only requires each point to satisfy the requirements indexed below it. [/example] The example shows why normality is the form of completeness used in practice: diagonal intersections preserve the kind of coherent measure-one information that ordinary intersections cannot capture at length $\kappa$. The possible obstruction is that a measurable cardinal is only defined using a nonprincipal $\kappa$-complete ultrafilter, and such a measure need not visibly have this diagonal closure. The theorem below removes that gap by showing that measurability supplies a normal measure after passing to a suitable derived measure. [quotetheorem:7409] [citeproof:7409] This result justifies treating normal measures as part of the basic measurable-cardinal toolkit, but it is not a statement that every measure is already normal. The construction changes the measure by pushing it forward along a carefully chosen function, and the well-foundedness of the ultrapower comparison is the hidden ingredient that makes the choice legitimate. Countable completeness is essential in that step: a countably incomplete nonprincipal ultrafilter on $\omega$ has an ill-founded ultrapower of the ordinals, so the minimal representative used above need not exist. Nonprincipality is also essential; the principal ultrafilter at $\gamma<\kappa$ is $\kappa$-complete, but its pushforwards remain principal and cannot become a normal measure in the large-cardinal sense. The theorem also identifies what normalisation does not repair. If an ultrafilter lacks enough completeness, it may split a countable decreasing sequence of large sets with empty intersection, and diagonal arguments then have no stable measure-one core. If the ultrafilter is principal, every regressive-function statement reduces to behaviour at one point and carries no information about $\kappa$. Once a normal measure is available, many arguments become parallel to stationary-set arguments, with measure-one sets replacing stationary sets and diagonal intersections replacing club intersections. ## Measurable Cardinals and First Consequences The central question is what kind of cardinal can support a nonprincipal $\kappa$-complete ultrafilter. The answer begins with measurability: a single measure already forces the cardinal far beyond the weakly compact and Mahlo phenomena from the previous chapters. [definition: Measurable Cardinal] An uncountable cardinal $\kappa$ is measurable if there exists a nonprincipal $\kappa$-complete ultrafilter on $\kappa$. [/definition] By the preceding theorem, this definition could equivalently ask for a normal measure. The first test of its strength is cofinality: a singular cardinal can be built from fewer than $\kappa$ bounded pieces, while a measure cannot regard such a union as large. [quotetheorem:7410] [citeproof:7410] Regularity says that $\kappa$ cannot be approached by a short cofinal sequence. Each hypothesis is doing work. A principal ultrafilter on a singular cardinal such as $\aleph_\omega$ is $\aleph_\omega$-complete in the formal intersection sense, but it concentrates on one point and therefore says nothing about the cofinal sequence witnessing singularity. A nonprincipal ultrafilter without $\kappa$-completeness also cannot prove regularity: on a singular $\kappa$ an ultrafilter extending the cofinal filter may treat every tail as large while failing to make the union of fewer than $\kappa$ null bounded pieces null. The proof above uses exactly the missing step, namely closure under fewer than $\kappa$ intersections. The conclusion has a sharp limitation. Regular cardinals need not be measurable, and regularity alone gives no power-set control; for instance $\aleph_1$ is regular in ZFC but need not be a strong limit. Thus the theorem is the first cofinality consequence of measurability, not a characterisation. The next obstruction concerns size: if some $\mathcal P(\lambda)$ for $\lambda<\kappa$ had size at least $\kappa$, the measure would decide too many binary coordinates at once. [quotetheorem:7411] [citeproof:7411] The strong limit theorem is the point where normality gives a power-set consequence rather than only a cofinality consequence. Nonprincipality is needed because a principal ultrafilter on any non-strong-limit cardinal, such as $\aleph_1$ in a model where $2^{\aleph_0}\ge\aleph_1$, satisfies all statements supported by its distinguished point and cannot constrain $2^\lambda$. Completeness is needed because the proof decides $\lambda<\kappa$ many coordinate sets at once; without $\kappa$-completeness, the simultaneous agreement set can disappear. Normality is the final strengthening: without it, the coordinate-disagreement function need not become constant on a measure-one set, so the proof has no fixed coordinate at which to contradict the definition of $X$. The conclusion is still only the first size threshold. Strong limitness alone does not imply regularity, as singular strong limit cardinals are consistent relative to ordinary set-theoretic hypotheses; regularity alone does not imply strong limitness, as $\aleph_1$ can fail the continuum bound. Under the definition used here, a measurable cardinal is an uncountable cardinal carrying a nonprincipal $\kappa$-complete ultrafilter, and normality is available by the normal-measure theorem above. Thus the strong-limit conclusion is a standard consequence of measurability, not an added convention. The next theorem packages the two independent measurable-cardinal consequences as inaccessibility, making explicit that measurable cardinals extend the hierarchy developed in the opening chapter. [quotetheorem:7412] [citeproof:7412] Inaccessibility is a packaging theorem rather than a new use of the measure, but the packaging matters because it locates measurability inside the earlier hierarchy. The assumptions cannot be weakened to a principal ultrafilter, and they cannot be replaced by regularity alone or strong limitness alone, since each property misses one half of strong inaccessibility. The theorem also does not yet assert that there are many inaccessible cardinals below $\kappa$; it only says that $\kappa$ itself is inaccessible. To state the next strengthening, we recall the earlier notion that a cardinal sees many inaccessible cardinals below it, measured by stationarity rather than mere unboundedness. [definition: Mahlo Cardinal] A strongly inaccessible cardinal $\kappa$ is Mahlo if the set of strongly inaccessible cardinals below $\kappa$ is stationary in $\kappa$. [/definition] For measurable cardinals, the measure gives a stronger conclusion than stationarity. The inaccessible cardinals below $\kappa$ form a measure-one set for any normal measure on $\kappa$, so every club must meet them. [quotetheorem:7413] [citeproof:7413] The pattern is important: to prove that many smaller ordinals have a property, assume the bad set is large and use a regressive witness to the failure. Normality then makes the witness constant on a large set, converting a varying local obstruction into a fixed bounded obstruction. Nonprincipality is needed because a principal ultrafilter can concentrate on a single non-inaccessible ordinal below a large $\kappa$ and would falsely treat that singleton as the only relevant large set. Completeness is needed in the regular-cardinal part when the proof intersects the sets $S_i$ for $i<\lambda$; without it, the cofinal sequences may be bounded at each coordinate on large sets while having no common large set on which all coordinates are bounded. Normality is needed in both halves because the functions $\alpha\mapsto\operatorname{cf}(\alpha)$ and $\alpha\mapsto\lambda(\alpha)$ are useful only after a constant large fibre is forced. The theorem is stronger than ordinary Mahloness because it does not merely say every club meets the inaccessibles; it says the inaccessibles are large for every chosen normal measure. Its limitation is equally important: it does not say that every inaccessible is measurable, nor that the inaccessible cardinals below $\kappa$ carry measures of their own. It says that the original measure on $\kappa$ concentrates on ordinals satisfying the earlier inaccessibility criteria. [example: Regressive Functions Constant on a Measure-One Set] Let $U$ be a normal measure on $\kappa$ and let $S\in U$. Write \begin{align*} R=\{\xi\in S:\xi \text{ is singular and } \xi\ne 0\}. \end{align*} If $R\in U$, define $f:R\to\kappa$ by $f(\xi)=\operatorname{cf}(\xi)$. For every $\xi\in R$, singularity means \begin{align*} \operatorname{cf}(\xi)<\xi. \end{align*} Thus \begin{align*} f(\xi)=\operatorname{cf}(\xi)<\xi \end{align*} for every $\xi\in R$, so $f$ is regressive on the measure-one set $R$. By the regressive-function form of normality, there is some $\lambda<\kappa$ such that \begin{align*} f^{-1}(\{\lambda\})\in U. \end{align*} Unwinding the definition of $f$ gives \begin{align*} f^{-1}(\{\lambda\})=\{\xi\in R:\operatorname{cf}(\xi)=\lambda\}. \end{align*} Since $R\subseteq S$, this is the same measure-one conclusion as saying that measure-one many singular ordinals in $S$ have the fixed cofinality $\lambda$. This is the constant-fibre mechanism used in the proof that normal measures cannot concentrate on singular ordinals, and hence must concentrate on regular and eventually inaccessible ordinals. [/example] The remaining section isolates the Ulam-matrix style of reasoning, a flexible way to build normality and partition contradictions from two-dimensional arrays of sets. ## Ulam Matrices and Normality Arguments Normality arguments often need to compare two incompatible ways of slicing $\kappa$. A Ulam matrix arranges subsets of $\kappa$ in rows and columns so that each row covers a large set while each column is small, or vice versa. The contradiction comes from asking an ultrafilter to choose one side of every partition. [definition: Ulam Matrix] Let $\kappa$ be an uncountable cardinal. An Ulam matrix on $\kappa$ is a function \begin{align*} A:\kappa\times\kappa\to\mathcal P(\kappa), \qquad A(\alpha,\beta)=A_{\alpha,\beta}, \end{align*} such that: 1. for each fixed $\alpha<\kappa$, the family $(A_{\alpha,\beta})_{\beta<\kappa}$ is a pairwise disjoint partition of $\kappa\setminus(\alpha+1)$; 2. for each fixed $\beta<\kappa$, the column set is \begin{align*} C_\beta=\bigcup_{\alpha<\kappa}A_{\alpha,\beta}; \end{align*} 3. the rows are indexed so that every point $\xi\in A_{\alpha,\beta}$ satisfies $\alpha<\xi$. [/definition] This definition is now a concrete bookkeeping device rather than a slogan. The row condition produces regressive partitions, because the row index is always below the point being partitioned. The column sets collect the same choices vertically, which lets an ultrafilter argument compare what happens row-by-row with what happens after a single column is selected. Different applications may build the entries in different ways, but the partition and regressiveness conditions are the data needed for theorem-level use. [quotetheorem:7414] [citeproof:7414] This measure version of Fodor's lemma is stronger than the stationary-set version when a normal measure is present: the constant set is not merely stationary, but measure-one. The theorem requires normality, not just $\kappa$-completeness; a $\kappa$-length family of null fibres can evade ordinary completeness unless diagonal closure is available. It also requires regressiveness, since an arbitrary function $S\to\kappa$ can split a measure-one set according to non-regressive information. It is the main engine behind normality arguments throughout the rest of the course, especially when a failure of reflection can be encoded by choosing a least witness below each ordinal. [example: A Normality Partition Argument] Let $U$ be a normal measure on $\kappa$, and let $S\in U$ be partitioned as \begin{align*} S=\bigcup_{\alpha<\kappa}S_\alpha \end{align*} with $S_\alpha\cap S_\beta=\varnothing$ whenever $\alpha\ne\beta$, and with $S_\alpha\subseteq\{\xi<\kappa:\alpha<\xi\}$ for every $\alpha<\kappa$. Since the pieces are pairwise disjoint and cover $S$, each $\xi\in S$ belongs to a unique piece $S_\alpha$. Define $f:S\to\kappa$ by setting $f(\xi)=\alpha$ for this unique $\alpha$. If $\xi\in S$ and $f(\xi)=\alpha$, then $\xi\in S_\alpha$. By the hypothesis on $S_\alpha$, \begin{align*} \xi\in S_\alpha\subseteq\{\eta<\kappa:\alpha<\eta\}. \end{align*} Therefore $\alpha<\xi$, and hence \begin{align*} f(\xi)=\alpha<\xi. \end{align*} Thus $f$ is regressive on $S$. By the normal-measure Fodor lemma above, there is some $\alpha_0<\kappa$ such that \begin{align*} f^{-1}(\{\alpha_0\})\in U. \end{align*} For $\xi\in S$, the definition of $f$ and the disjointness of the partition give \begin{align*} \xi\in f^{-1}(\{\alpha_0\}) \text{ exactly when } \xi\in S_{\alpha_0}. \end{align*} Hence \begin{align*} f^{-1}(\{\alpha_0\})=S_{\alpha_0}. \end{align*} So $S_{\alpha_0}\in U$. A normal measure therefore cannot divide a measure-one set into a regressive partition whose every piece is null; one piece must already be measure-one. [/example] The partition example is the one-dimensional form of the Ulam idea: a measure must select a large cell when the indexing is regressive. The next theorem records the normalisation principle in matrix language, where the rows represent many such partitions simultaneously and the columns encode the attempted failures of normality. [quotetheorem:7415] [citeproof:7415] The matrix principle is not a new large-cardinal existence theorem; the existence theorem was the earlier normal-measure construction. Its role is organizational: it packages regressive-function normality into a two-dimensional bookkeeping form that is convenient when many partitions are being compared simultaneously. The regressiveness of the selected column is essential, because without the inequality $\beta(\alpha)<\alpha$ Fodor's lemma cannot force a constant column. The chapter's main conceptual shift is that largeness at $\kappa$ is no longer expressed only through clubs, stationary sets, or reflection statements. A measurable cardinal carries a complete ultrafilter that behaves like a two-valued probability measure closed under all smaller intersections. Normality makes this measure compatible with the ordinal structure below $\kappa$, and that compatibility is what turns measurability into inaccessibility, Mahloness, and a template for the elementary embeddings studied next. Measurability first appeared as the existence of a highly complete, normal way to decide which subsets of $\kappa$ are large. The next chapter explains why this is not merely a measure-theoretic notion: every such ultrafilter produces an elementary embedding, and the embedding remembers the large-cardinal strength. # 4. Ultrapowers and Elementary Embeddings This chapter changes the viewpoint from filters on a cardinal to elementary embeddings of the universe. The guiding question is how an ultrafilter can produce a new universe-like structure, and how the resulting map remembers the size and completeness of the original measure. The central bridge is the ultrapower construction: functions modulo an ultrafilter become points of a model, and Los's theorem makes first-order truth pass through the quotient. The previous chapters built large cardinals through closure, reflection, and combinatorial strength. Measurable cardinals enter at a different level: they are cardinals that carry enough measure-theoretic structure to move the universe by an elementary embedding. This chapter develops both directions of that equivalence. ## Building Ultrapowers from Complete Ultrafilters The first problem is how to turn an ultrafilter into a structure that still satisfies the same first-order statements as the ambient universe. A function $f:I \to V$ should be thought of as an $I$-indexed approximation to a single object, and the ultrafilter decides when two such approximations represent the same point. [definition: Ultrapower Equivalence] Let $U$ be an ultrafilter on a set $I$. For functions $f,g:I\to V$, define \begin{align*} f =_U g \quad \text{iff} \quad \{i\in I : f(i)=g(i)\}\in U. \end{align*} The equivalence class of $f$ is denoted $[f]_U$. [/definition] This quotient relation discards information on a $U$-small set, but equality alone does not yet give a model of set theory. We need a membership relation on equivalence classes so that formulas involving $\in$ have a coordinatewise interpretation. [definition: Ultrapower of the Universe] Let $U$ be an ultrafilter on a set $I$. The ultrapower $\operatorname{Ult}(V,U)$ is the class of equivalence classes $[f]_U$ for functions $f:I\to V$, equipped with the relation \begin{align*} [f]_U \in_U [g]_U \quad \text{iff} \quad \{i\in I : f(i)\in g(i)\}\in U. \end{align*} The diagonal embedding $j_U:V\to \operatorname{Ult}(V,U)$ is defined by $j_U(a)=[c_a]_U$, where $c_a(i)=a$ for all $i\in I$. [/definition] The relation $\in_U$ is well-defined because the ultrafilter respects equality on a large set. The next problem is to prove that every first-order formula, not only atomic equality and membership, is evaluated by checking where it holds on a $U$-large set. [quotetheorem:7416] [citeproof:7416] Los's theorem says that the ultrapower is not merely a quotient class: it is elementarily related to $V$. The ultrafilter hypothesis is doing real work here. For a proper filter which is not an ultrafilter, a formula and its negation need not be decided modulo the filter, so Boolean negation can fail in the quotient semantics. Thus the theorem is not a general statement about all reduced products: maximality is what makes first-order truth two-valued after quotienting. There is also an important limitation. Los's theorem gives elementary agreement between $V$ and the quotient structure in the language using $\in_U$, but it does not say that the quotient is well-founded or that it is already a transitive class. Those are separate set-theoretic properties of the quotient relation, and they are the reason completeness hypotheses enter next. We first extract the elementary embedding given by constant functions, then return to the well-foundedness obstruction. [quotetheorem:4289] [citeproof:4289] The diagonal embedding is elementary because constants turn every ground-model truth value into either the whole index set or the empty set. This argument also shows why a proper ultrafilter is needed in the background: if the quotient semantics were built from a non-maximal filter, the inductive step for negation could leave a formula undecided, and the constant-map argument would no longer sit inside a fully elementary quotient. Elementarity of $j_U$ also does not imply that the quotient has ordinary membership as its relation. A non-countably-complete ultrafilter may still produce an elementary quotient with an infinite descending $\in_U$-chain, so large cardinal applications require an additional closure assumption. The needed extra hypothesis is closure of the ultrafilter under long intersections. We now name the precise completeness property that prevents descending chains. [definition: Kappa Complete Ultrafilter] Let $\kappa$ be an uncountable cardinal. An ultrafilter $U$ on a set $I$ is $\kappa$-complete if for every family $(A_\alpha)_{\alpha<\lambda}$ of members of $U$ with $\lambda<\kappa$, \begin{align*} \bigcap_{\alpha<\lambda} A_\alpha \in U. \end{align*} [/definition] For the first applications, countable completeness is the minimum requirement because ill-foundedness is witnessed by a descending $\omega$-sequence. We next need the exact theorem converting countable completeness into well-foundedness of the ultrapower. [example: Constant and Identity Functions] Let $U$ be a uniform $\kappa$-complete non-principal ultrafilter on the cardinal $\kappa$, and form $\operatorname{Ult}(V,U)$. For each $\alpha<\kappa$, the diagonal embedding sends $\alpha$ to the class of the constant function: \begin{align*} j_U(\alpha)=[c_\alpha]_U,\qquad c_\alpha(\xi)=\alpha. \end{align*} The identity function $\operatorname{id}_\kappa(\xi)=\xi$ gives the class $[\operatorname{id}_\kappa]_U$. To compare these two classes in the ultrapower ordering on ordinals, compute the set of coordinates where the comparison holds: \begin{align*} \{\xi<\kappa:c_\alpha(\xi)<\operatorname{id}_\kappa(\xi)\}=\{\xi<\kappa:\alpha<\xi\}. \end{align*} The complement of this set is $\{\xi<\kappa:\xi\leq \alpha\}=\alpha+1$, which has cardinality $<\kappa$. Since $U$ is non-principal, no singleton belongs to $U$; since $U$ is $\kappa$-complete, the union of fewer than $\kappa$ many singletons cannot belong to $U$. Hence $\alpha+1\notin U$, so by maximality of the ultrafilter, \begin{align*} \{\xi<\kappa:\alpha<\xi\}\in U. \end{align*} Therefore $[c_\alpha]_U<_U[\operatorname{id}_\kappa]_U$, that is, \begin{align*} j_U(\alpha)<_U[\operatorname{id}_\kappa]_U. \end{align*} Thus the identity class lies above every embedded ordinal below $\kappa$; after the Mostowski collapse in the normal-measure presentation, this class is the point usually denoted by $\kappa$ in the target model. [/example] This example is the computational heart of ultrapowers by measures on $\kappa$: ordinary functions $f:\kappa\to V$ become values of $j(f)$ at the special point represented by the identity. The remaining problem is to justify that this computation takes place inside a transitive target model. ## Well-Foundedness and the Mostowski Collapse The ultrapower relation $\in_U$ is defined syntactically, but large cardinal embeddings need a genuine transitive class as target. The next question is therefore when the quotient structure has no infinite descending membership chains, so that it can be collapsed to a transitive model. [definition: Well-Founded Extensional Relation] A binary relation $E$ on a class $A$ is well-founded if every non-empty subclass $B\subset A$ has an $E$-minimal element. It is extensional if for all $a,b\in A$, \begin{align*} (\forall x\in A\, (xEa \iff xEb)) \implies a=b. \end{align*} [/definition] Extensionality gives uniqueness of a transitive representative, while well-foundedness makes recursive rank assignment possible. For ultrapowers, extensionality follows from equality modulo $U$, so the key problem is to prove well-foundedness from countable completeness. [quotetheorem:7417] [citeproof:7417] This theorem transforms an abstract quotient into a structure eligible for collapse, but its hypothesis cannot be ignored. If an ultrafilter is not countably complete, the coordinatewise witnesses to a descending chain may live on a decreasing sequence of large sets whose total intersection is empty, so there need not be any single index at which foundation in $V$ detects the contradiction. This is the standard mechanism by which ultrapowers by non-countably-complete ultrafilters can be ill-founded. Well-foundedness also does not by itself make the quotient transitive. The objects are still equivalence classes of functions and the relation is still $\in_U$, so the structure must be converted into one whose membership relation is the ambient $\in$. We need the Mostowski collapse theorem to replace the quotient relation by ordinary membership in a transitive class. [quotetheorem:4842] [citeproof:4842] The collapse requires both hypotheses. Without well-foundedness, the recursive definition of $\pi(a)$ may chase an infinite descending chain and fail to assign sets by rank. Without extensionality, two distinct elements with the same predecessors would be collapsed to the same set, so the resulting map would not be an isomorphism of the original relation. The theorem also does not create elementarity from nothing: it preserves the first-order information already present in the well-founded extensional structure by transporting the relation through an isomorphism. After the collapse, we usually identify the ultrapower with its transitive collapse and write $j:V\to M$. This convention hides the quotient notation, so we need a definition that records exactly which map is being used. [definition: Ultrapower Embedding] Let $U$ be a countably complete ultrafilter on $I$. The ultrapower embedding derived from $U$ is the composition \begin{align*} j:V\xrightarrow{j_U}\operatorname{Ult}(V,U)\xrightarrow{\pi}M, \end{align*} where $\pi$ is the Mostowski collapse and $M$ is transitive. [/definition] The target $M$ is an inner-model-like class: it is transitive and satisfies every first-order sentence of ZFC because $V$ does and $j$ is elementary. The quotient representation remains useful because it lets us compute values of the embedding inside $M$. [example: Computing an Ultrapower Value] Let $U$ be a countably complete ultrafilter on $\kappa$, let $\pi:\operatorname{Ult}(V,U)\to M$ be the Mostowski collapse, and let $j=\pi\circ j_U$. For a function $f:\kappa\to V$, the object $j(f)$ is \begin{align*} j(f)=\pi(j_U(f))=\pi([c_f]_U), \end{align*} where $c_f(\xi)=f$ for every $\xi<\kappa$. The special argument represented by the identity function is $\pi([\operatorname{id}_\kappa]_U)$, with $\operatorname{id}_\kappa(\xi)=\xi$. Function application in the ultrapower is computed coordinatewise: the value of $[c_f]_U$ at $[\operatorname{id}_\kappa]_U$ is the class of the function $\xi\mapsto c_f(\xi)(\operatorname{id}_\kappa(\xi))$. For each $\xi<\kappa$, \begin{align*} c_f(\xi)(\operatorname{id}_\kappa(\xi))=f(\xi). \end{align*} Thus this coordinatewise value is exactly the function $f$, so in the ultrapower presentation, \begin{align*} [c_f]_U([\operatorname{id}_\kappa]_U)=[f]_U. \end{align*} Applying the collapse map, which preserves the interpreted membership and function relations, gives \begin{align*} j(f)(\pi([\operatorname{id}_\kappa]_U))=\pi([f]_U). \end{align*} For example, if $f(\xi)=\xi+1$, then the same computation gives \begin{align*} j(f)(\pi([\operatorname{id}_\kappa]_U))=\pi([\xi\mapsto \xi+1]_U). \end{align*} Moreover, \begin{align*} \{\xi<\kappa:\operatorname{id}_\kappa(\xi)<f(\xi)\}=\{\xi<\kappa:\xi<\xi+1\}=\kappa\in U, \end{align*} so $[\operatorname{id}_\kappa]_U<_U[f]_U$. If an ordinal-valued $g$ satisfied $[\operatorname{id}_\kappa]_U<_U[g]_U<_U[f]_U$, then \begin{align*} \{\xi<\kappa:\xi<g(\xi)<\xi+1\}\in U, \end{align*} but no ordinal lies strictly between $\xi$ and $\xi+1$, so this set is empty, contradicting $\varnothing\notin U$. Therefore $\pi([\xi\mapsto \xi+1]_U)$ is the successor of $\pi([\operatorname{id}_\kappa]_U)$ in the collapsed ultrapower ordering. [/example] This calculation explains why ultrapowers are concrete despite their high-level appearance: every element of the ultrapower target is represented by some function in the ground universe. The next step is to identify the first ordinal at which the embedding can differ from the identity. ## Critical Points of Elementary Embeddings Once the ultrapower has been collapsed, the next question is how to measure the first non-identity behaviour of an elementary embedding. If $j:V\to M$ is elementary and nontrivial, it cannot move a small set before it moves some ordinal, because sets are built by rank from earlier sets. [definition: Critical Point] Let $M$ be a transitive class and let $j:V\to M$ be a non-identity elementary embedding. The critical point of $j$ is \begin{align*} \operatorname{crit}(j)=\min\{\alpha\in\operatorname{Ord}:j(\alpha)\ne \alpha\}. \end{align*} [/definition] The critical point is the exact location where the embedding starts to see new structure. We need its basic rank-theoretic properties because they explain why a measure on $\kappa$ yields an embedding that fixes all of $V_\kappa$ but moves $\kappa$ itself. The central issue is that ordinals below the critical point determine all sets of rank below it. [quotetheorem:7418] [citeproof:7418] These properties show that the critical point is a genuine threshold for the embedding. Transitivity of $M$ is essential in the rank argument: it lets membership in $j(x)$ be read as ordinary membership in a class whose elements have already been compared with their ground-model preimages. Without a transitive target, the statement $j(x)=\{j(y):y\in x\}$ would not give the same rank-by-rank control over actual elements. The theorem is also deliberately qualitative. It says that everything below $V_\kappa$ is fixed and that $\kappa$ moves upward, but it does not identify $j(\kappa)$, the closure properties of $M$, or which functions represent elements of the target in a given ultrapower. For an ultrapower by a measure on $\kappa$, we now need to prove that the threshold is exactly $\kappa$, rather than some smaller ordinal. [quotetheorem:7419] [citeproof:7419] This theorem also gives a useful normal form: small-valued functions collapse to constants, while functions with range cofinal in $\kappa$ can witness that $j(\kappa)$ has moved. Both hypotheses have visible roles. If $\kappa$-completeness is removed, a map $f:\kappa\to\alpha$ for $\alpha<\kappa$ can have all fibres outside $U$, producing a genuinely new ordinal below $\kappa$ in the quotient. If non-principality is removed, a principal ultrafilter gives a quotient isomorphic to $V$ and the associated embedding has no moved ordinal at all. This theorem identifies the first moved ordinal only. It does not assert that the ultrafilter is normal, does not compute $j(\kappa)$ beyond showing $j(\kappa)>\kappa$, and does not give closure properties such as $M^\lambda\subset M$. The fibre computation is important enough to record as an example. [example: Small-Valued Functions Become Constants] Let $U$ be a $\kappa$-complete ultrafilter on $\kappa$, and let $f:\kappa\to\alpha$ with $\alpha<\kappa$. For each $\beta<\alpha$, set \begin{align*} A_\beta=\{\xi<\kappa:f(\xi)=\beta\}. \end{align*} These fibres cover $\kappa$ because every value $f(\xi)$ is some ordinal below $\alpha$, and they are pairwise disjoint because $f(\xi)$ has only one value. At least one fibre belongs to $U$. If no $A_\beta$ belonged to $U$, then since $U$ is an ultrafilter, each complement $\kappa\setminus A_\beta$ would belong to $U$. Because there are $\alpha<\kappa$ many such complements and $U$ is $\kappa$-complete, \begin{align*} \bigcap_{\beta<\alpha}(\kappa\setminus A_\beta)\in U. \end{align*} But the fibres cover $\kappa$, so no $\xi<\kappa$ lies outside all of them, hence \begin{align*} \bigcap_{\beta<\alpha}(\kappa\setminus A_\beta)=\varnothing. \end{align*} This contradicts that an ultrafilter is proper, so some $A_\beta$ is in $U$. There is at most one such fibre: if $A_\beta\in U$ and $A_\gamma\in U$ with $\beta\ne\gamma$, then closure of $U$ under finite intersections gives \begin{align*} A_\beta\cap A_\gamma\in U. \end{align*} Since the fibres are disjoint, $A_\beta\cap A_\gamma=\varnothing$, again impossible for a proper ultrafilter. Therefore there is a unique $\beta<\alpha$ such that $A_\beta\in U$. For this $\beta$, \begin{align*} \{\xi<\kappa:f(\xi)=c_\beta(\xi)\}=\{\xi<\kappa:f(\xi)=\beta\}=A_\beta\in U. \end{align*} By the definition of ultrapower equivalence, $[f]_U=[c_\beta]_U$. Thus every function with range below $\alpha<\kappa$ represents a constant ordinal in the ultrapower, so it creates no new ordinal below $\kappa$. [/example] This fibre argument is repeatedly used in large cardinal theory. It is the local reason that a $\kappa$-complete measure has critical point at least $\kappa$, and it prepares the reverse construction from embeddings back to measures. ## Deriving Measures from Embeddings The ultrapower construction sends a measure to an elementary embedding. The reverse problem is whether an elementary embedding with critical point $\kappa$ contains a measure on $\kappa$. The answer is yes: the embedding tells us which subsets of $\kappa$ contain the distinguished point $\kappa$ after applying $j$. [definition: Measure Derived from an Embedding] Let $j:V\to M$ be an elementary embedding with transitive target $M$, and let $\kappa=\operatorname{crit}(j)$. Define \begin{align*} U_j=\{X\subset\kappa: \kappa\in j(X)\}. \end{align*} [/definition] This definition is the inverse of the ultrapower calculation $j(f)(\kappa)=[f]_U$ in the normal setting. We need to verify that it really gives a measure: it must decide every subset of $\kappa$, avoid singletons, and be closed under intersections of length below $\kappa$. [quotetheorem:7420] [citeproof:7420] Thus an elementary embedding with critical point $\kappa$ automatically packages a measure on $\kappa$, and each assumption in the construction has a specific purpose. Elementarity is needed to transport complements and intersections through $j$; without it, the rule $X\mapsto \kappa\in j(X)$ need not decide complements or respect intersections. The condition $\operatorname{crit}(j)=\kappa$ is what fixes all smaller ordinals and all shorter indexing sequences, which is why singletons are omitted and intersections of length $<\kappa$ remain large. Transitivity of $M$ is also part of the bookkeeping, since the membership test $\kappa\in j(X)$ is an ordinary membership statement in the target. If $\kappa$ were merely a moved ordinal but not the critical point, the same definition could fail to be non-principal or fail to have the intended completeness. We need a name for cardinals carrying such measures so that the measure formulation and embedding formulation can be stated as one equivalence. [definition: Measurable Cardinal] An uncountable cardinal $\kappa$ is measurable if there exists a non-principal $\kappa$-complete ultrafilter on $\kappa$. [/definition] The embedding formulation is often the working definition in modern large cardinal theory, but at this point there are two apparently different notions on the table. A measure produces an ultrapower embedding only if the ultrapower is well-founded, while an arbitrary elementary embedding yields a measure only if the derived membership test really has the required completeness and non-principality. The theorem below resolves this possible mismatch by identifying exactly when the measure and embedding formulations express the same large-cardinal strength. [quotetheorem:7421] [citeproof:7421] This equivalence is the first major instance of the elementary-embedding paradigm, and the hypotheses should be read carefully. Non-principality corresponds to nontriviality: a principal measure gives only the evaluation embedding, while a nontrivial elementary embedding has a genuine critical point. $\kappa$-completeness is what supplies well-foundedness of the ultrapower in the forward direction and closure under short intersections in the reverse direction. Transitivity of $M$ is what lets the derived-measure test use ordinary membership rather than an external quotient relation. The theorem does not say that every elementary embedding with critical point $\kappa$ is the ultrapower embedding by its derived measure, nor does it assert normality of the measure. Those stronger conclusions require additional hypotheses or further analysis of the comparison map from an ultrapower into $M$. Later large cardinals refine the same pattern by asking how much of $V$ the target $M$ contains, how closed $M$ is under sequences, and how far above the critical point the embedding sees. [example: The Derived Measure in an Ultrapower] Suppose $U$ is a normal measure on $\kappa$ and $j:V\to M$ is the ultrapower embedding, with the collapsed point $\kappa$ identified with the class $[\operatorname{id}_\kappa]_U$. Let $X\subset\kappa$. In the ultrapower presentation, $j(X)$ is represented by the constant function $c_X:\kappa\to V$ with $c_X(\xi)=X$ for every $\xi<\kappa$, so \begin{align*} j_U(X)=[c_X]_U. \end{align*} Under the usual identification of the collapsed ultrapower with $M$, the membership test $\kappa\in j(X)$ is therefore the same as \begin{align*} [\operatorname{id}_\kappa]_U\in_U[c_X]_U. \end{align*} By the definition of ultrapower membership, \begin{align*} [\operatorname{id}_\kappa]_U\in_U[c_X]_U \quad \text{iff} \quad \{\xi<\kappa:\operatorname{id}_\kappa(\xi)\in c_X(\xi)\}\in U. \end{align*} For each $\xi<\kappa$, $\operatorname{id}_\kappa(\xi)=\xi$ and $c_X(\xi)=X$, hence \begin{align*} \{\xi<\kappa:\operatorname{id}_\kappa(\xi)\in c_X(\xi)\}=\{\xi<\kappa:\xi\in X\}. \end{align*} Since $X\subset\kappa$, \begin{align*} \{\xi<\kappa:\xi\in X\}=X. \end{align*} Combining the equivalences gives \begin{align*} \kappa\in j(X) \quad \text{iff} \quad X\in U. \end{align*} Thus the measure derived from the ultrapower embedding, \begin{align*} \{X\subset\kappa:\kappa\in j(X)\}, \end{align*} is exactly the original ultrafilter $U$. [/example] The chapter closes the circle: measures produce ultrapowers, well-founded ultrapowers collapse to transitive targets, and elementary embeddings with critical point $\kappa$ recover measures. Chapter 5 uses this circle as the template for stronger embedding axioms, where the target model is required to contain longer rank segments or to be closed under longer sequences. Ultrapowers reveal measurability as the first instance of a general embedding pattern. The next chapter strengthens that pattern by requiring the target model to see more of the universe, turning closure and rank containment into new large-cardinal axioms. # 5. Stronger Embedding Axioms The previous chapters developed large cardinals from reflection and elementary embeddings, with measurability as the first point where an ultrafilter produces a nonidentity map $j:V\to M$. This chapter strengthens the embedding viewpoint by asking how much of the universe the target model $M$ must contain. Strong cardinals demand that $M$ contain long initial segments $V_\lambda$, while supercompact cardinals demand closure under long sequences and are equivalently described by fine normal measures on $P_\kappa(\lambda)$. The organizing question is: what extra mathematical content is gained by asking an elementary embedding to preserve not only truth, but also enough sets and enough sequences to reconstruct large pieces of the universe inside its target? ## Initial Segments Inside The Target Model Measurability gives an embedding $j:V\to M$ with critical point $\kappa$, but the ultrapower target need not contain every set of a prescribed rank. The next strengthening asks for embeddings whose target models know a chosen initial segment of the cumulative hierarchy. [definition: Lambda Strong Cardinal] Let $\kappa$ be a cardinal and let $\lambda$ be an ordinal with $\lambda\geq\kappa$. The cardinal $\kappa$ is $\lambda$-strong if there is an elementary embedding $j:V\to M$ such that $M$ is transitive, $\operatorname{crit}(j)=\kappa$, $j(\kappa)>\lambda$, and $V_\lambda\subset M$. [/definition] The point of $V_\lambda\subset M$ is that the target is not merely a quotient-like image of the universe; it literally contains every set built before stage $\lambda$. Since a single value of $\lambda$ gives only a local version of this phenomenon, the natural next question is whether one cardinal can supply such embeddings for all ranks. [definition: Strong Cardinal] A cardinal $\kappa$ is strong if $\kappa$ is $\lambda$-strong for every ordinal $\lambda$. [/definition] Strongness is an embedding axiom about rank-initial approximation. It does not require $M$ to be closed under all sequences of a given length, so the next useful test is to inspect what rank containment lets us read from a fixed embedding. [example: Reading Rank Containment From An Embedding] Suppose $j:V\to M$ has critical point $\kappa$, $j(\kappa)>\lambda$, and $V_\lambda\subset M$. We show that for every $\alpha<\lambda$, every subset of $V_\alpha$ belonging to $V$ is also an element of $M$. Fix $\alpha<\lambda$ and let $A\subset V_\alpha$ with $A\in V$. By the successor step in the cumulative hierarchy, \begin{align*} V_{\alpha+1}=P(V_\alpha). \end{align*} Since $A\subset V_\alpha$, this means $A\in P(V_\alpha)$, hence \begin{align*} A\in V_{\alpha+1}. \end{align*} Because $\alpha<\lambda$, we have $\alpha+1\leq\lambda$, so monotonicity of the cumulative hierarchy gives \begin{align*} V_{\alpha+1}\subset V_\lambda. \end{align*} Combining the inclusions, \begin{align*} A\in V_{\alpha+1}\subset V_\lambda\subset M. \end{align*} Thus $P(V_\alpha)^V\subset M$ for every $\alpha<\lambda$. A $\lambda$-strong embedding therefore gives the target model access to the actual subsets of each lower rank, not merely to objects that appear as values of $j$. [/example] This example explains why strongness is useful in inner model theory: the target model can compare itself with $V$ on a long rank initial segment. The next step asks for a different kind of access, namely access to long sequences. ## Supercompactness And Closure Under Sequences The problem with rank containment is that it does not directly say that $M$ is closed under functions from $\lambda$ into $M$. Supercompactness strengthens the embedding axiom by requiring such closure, and this turns the embedding into a tool for controlling structures of size at most $\lambda$. [definition: Lambda Supercompact Cardinal] Let $\kappa$ be a cardinal and let $\lambda\geq\kappa$ be an ordinal. The cardinal $\kappa$ is $\lambda$-supercompact if there is an elementary embedding $j:V\to M$ such that $M$ is transitive, $\operatorname{crit}(j)=\kappa$, $j(\kappa)>\lambda$, and $M^\lambda\subset M$. [/definition] Here $M^\lambda\subset M$ means that every function $f:\lambda\to M$ which belongs to $V$ is itself an element of $M$. Since one length $\lambda$ gives only a local closure requirement, the next axiom asks for embeddings with this closure at every length. [definition: Supercompact Cardinal] A cardinal $\kappa$ is supercompact if $\kappa$ is $\lambda$-supercompact for every ordinal $\lambda\geq\kappa$. [/definition] Supercompactness is much stronger than measurability because it gives embeddings whose targets are closed under arbitrarily long sequences. The obstruction in using such embeddings is that each clause has to survive external interpretation: the target must be transitive, the critical point must be the intended cardinal, $j(\kappa)$ must lie beyond the length being measured, and the target must contain the relevant $\lambda$-sequences. In the rest of the course, the phrase "$\lambda$-supercompactness embedding" means exactly an embedding with these clauses. The embedding definition is compact, but each hypothesis has a different logical role. Transitivity is a necessity, not a convenience: if $M$ is replaced by an isomorphic copy with a transported membership relation rather than the real $\in$ relation, elementarity as an abstract structure can survive while statements such as "$x\in M$" and "$s:\lambda\to M$ belongs to $M$" no longer have their usual external meaning. The Mostowski collapse is precisely the operation that removes this failure model. The critical point requirement is equally specific. If $\mu<\kappa$ is supercompact and $j:V\to M$ has $\operatorname{crit}(j)=\mu$, then the same closure statement may hold for long $\lambda$, but the derived ultrafilter is on $P_\mu(\lambda)$ and witnesses largeness of $\mu$, not of $\kappa$. The inequality $j(\kappa)>\lambda$ cannot be dropped either. The derived measure construction below uses the seed $j``\lambda$ as an element of $P_{j(\kappa)}(j(\lambda))$; if $j(\kappa)\leq\lambda$, then the seed has size $\lambda$ and falls outside that domain. A concrete way to see the obstruction is to try the derivation with any embedding whose critical point is too small for the chosen length: the set $j``\lambda$ is no longer a small subset from the target's point of view, so the definition $A\in U$ iff $j``\lambda\in j(A)$ does not even type-check as a measure on $P_\kappa(\lambda)$. Finally, $M^\lambda\subset M$ is the hypothesis that separates supercompactness from rank containment. In relative consistency models with a strong cardinal that is not supercompact, a strongness embedding can contain $V_\lambda$ while there is no fine normal $\kappa$-complete measure on $P_\kappa(\lambda)$ for some $\lambda$; at that length, a $\lambda$-sequence through the target must be missing. What the theorem does not yet reveal is the ultrafilter structure from which calculations are usually made. To expose that structure, we must identify the correct domain for the measure associated with a length $\lambda$ supercompactness embedding. ## Fine Normal Measures On $P_\kappa(\lambda)$ For a measurable cardinal, the ultrafilter lives on $\kappa$. For supercompactness at length $\lambda$, the natural domain is $P_\kappa(\lambda)$, the set of subsets of $\lambda$ of size below $\kappa$. [definition: The Set $P_\kappa(\lambda)$] Let $\kappa$ be a cardinal and let $\lambda$ be an ordinal. The set $P_\kappa(\lambda)$ is \begin{align*} P_\kappa(\lambda)=\{x\subset \lambda: |x|<\kappa\}. \end{align*} [/definition] The set $P_\kappa(\lambda)$ records small approximations to $\lambda$. A measure on this set should regard the basic set of all small subsets containing a fixed point $\alpha<\lambda$ as large, which leads to the fineness condition. [definition: Fine Measure] Let $\kappa$ be a cardinal, let $\lambda$ be an ordinal, and let $U\subset P(P_\kappa(\lambda))$ be an ultrafilter on the full power set of $P_\kappa(\lambda)$. The ultrafilter $U$ is fine if for every $\alpha<\lambda$, \begin{align*} \{x\in P_\kappa(\lambda):\alpha\in x\}\in U. \end{align*} [/definition] Fineness says that the ultrapower sees each point of $\lambda$ represented by membership in almost every small set. To obtain the analogue of a normal measure on a measurable cardinal, we also need a pressing-down principle for functions that choose an element from each small set. [definition: Normal Measure on $P_\kappa(\lambda)$] Let $U\subset P(P_\kappa(\lambda))$ be a $\kappa$-complete ultrafilter on the full power set of $P_\kappa(\lambda)$. The ultrafilter $U$ is normal if whenever $f:P_\kappa(\lambda)\to\lambda$ has some $A\in U$ such that for every $x\in A$, $f(x)\in x$, there is some $\alpha<\lambda$ such that \begin{align*} \{x\in P_\kappa(\lambda):f(x)=\alpha\}\in U. \end{align*} [/definition] The normality condition is the analogue of pressing down on a stationary set. Together with fineness and $\kappa$-completeness, it raises the main question of this section: do these measure properties exactly recover the supercompactness embedding with $M^\lambda\subset M$? [quotetheorem:7422] [citeproof:7422] This theorem is the supercompact analogue of the measurable-cardinal ultrapower construction, and each extra adjective on the measure corresponds to a feature of the target embedding. The hypotheses are not interchangeable. If fineness is omitted, the principal ultrafilter concentrating on a fixed $x_0\in P_\kappa(\lambda)$ produces an ultrapower that sees only the coordinates in $x_0$; for any $\alpha\in\lambda\setminus x_0$, the set of small subsets containing $\alpha$ is not measure-one. If $\kappa$-completeness is omitted, countable intersections of measure-one requirements can already fail, and the resulting ultrapower may move small ordinals or fail to be well founded. If normality is omitted, a regressive map $f(x)\in x$ may vary on every measure-one set, so the ultrapower has no constant representative for a coordinate choice that the closure argument needs to identify. The hypothesis $\lambda\geq\kappa$ marks the genuinely supercompact range; below $\kappa$, the statement collapses toward ordinary completeness phenomena rather than long-sequence closure. The theorem does not say that every ultrafilter on $P_\kappa(\lambda)$ gives supercompactness, only that the fine, normal, $\kappa$-complete ones do. Since fineness is the most visible new condition, it is worth seeing what it does at the level of coordinates. [example: Fine Measure Coordinates] Let $U$ be a fine ultrafilter on $P_\kappa(\lambda)$, and let $c:P_\kappa(\lambda)\to P_\kappa(\lambda)$ be the coordinate map $c(x)=x$. Fix $\alpha<\lambda$. By fineness, \begin{align*} X_\alpha=\{x\in P_\kappa(\lambda):\alpha\in x\}\in U. \end{align*} In the ultrapower by $U$, the constant function with value $\alpha$ represents the element $j(\alpha)$. Since $c(x)=x$, the relation $j(\alpha)\in[c]_U$ is witnessed exactly by the $U$-large set \begin{align*} \{x\in P_\kappa(\lambda):\alpha\in c(x)\}=\{x\in P_\kappa(\lambda):\alpha\in x\}=X_\alpha. \end{align*} Thus the seed $[c]_U$ contains the ultrapower copy of each coordinate $\alpha<\lambda$. Now fix finitely many ordinals $\alpha_1,\dots,\alpha_n<\lambda$. For each $i$ with $1\leq i\leq n$, fineness gives \begin{align*} X_{\alpha_i}=\{x\in P_\kappa(\lambda):\alpha_i\in x\}\in U. \end{align*} Since $U$ is $\kappa$-complete, and $n<\kappa$, the finite intersection is still in $U$: \begin{align*} \bigcap_{i=1}^n X_{\alpha_i}\in U. \end{align*} Unwinding the definition of the intersection, \begin{align*} \bigcap_{i=1}^n X_{\alpha_i}=\{x\in P_\kappa(\lambda):\alpha_1\in x,\dots,\alpha_n\in x\}. \end{align*} So every finite pattern of prescribed coordinates below $\lambda$ appears on a $U$-large set of small subsets, which is why fineness lets the ultrapower recover the visible coordinates of $\lambda$ from the map $x\mapsto x$. [/example] The example shows why measures on $P_\kappa(\lambda)$ are suited to supercompactness: they encode $\lambda$ by looking at many small approximations at once. We can now compare the large cardinal notions seen so far. ## Measurable, Strong, And Supercompact Cardinals The hierarchy is not only a list of definitions; it records how much structure the embedding target retains. A measurable cardinal gives a nonidentity elementary embedding, a strong cardinal gives rank containment, and a supercompact cardinal gives closure under long sequences. [quotetheorem:7423] [citeproof:7423] This proof explains the intuitive reason for the implication: sufficiently long sequence closure is strong enough to reconstruct rank initial segments. The word sufficiently is important; closure under $\lambda$-sequences alone need not code all members of $V_\lambda$, because the [transitive closure](/theorems/1493) of a set of rank below $\lambda$ can have cardinality larger than $\lambda$. For example, when $\lambda$ is singular, sets of rank below $\lambda$ can have transitive closures whose sizes are cofinal in or above $\lambda$, so a code of length exactly $\lambda$ may not capture the object. The theorem therefore uses supercompactness at a larger scale and then descends to the desired rank. Without full supercompactness, this scaling step is unavailable: a single $\theta$-supercompactness embedding proves only the corresponding bounded strongness conclusion, and a merely strong embedding may contain $V_\lambda$ while omitting a $\lambda$-sequence through its target. Thus the theorem proves only one direction; rank containment is weaker than sequence closure, and the argument gives no way to recover $M^\lambda\subset M$ from $V_\lambda\subset M$. To place this result in the earlier part of the course, we next add measurability to the comparison. [quotetheorem:7424] [citeproof:7424] The hierarchy should be read as increasing expressive power of the embedding target, but the theorem itself proves only forward implications. A normal measure derived from a strongness embedding gives measurability, yet it does not remember the rank-containment property that produced it; the ordinary ultrapower by a normal measure on $\kappa$ has no reason to contain $V_{\kappa+2}$, let alone $V_\lambda$ for arbitrary $\lambda$. Similarly, the proof that supercompactness implies strongness deliberately throws away closure information after using it to obtain rank containment. There are relative consistency models, obtained by inner model and forcing methods, in which a measurable cardinal exists with no strong cardinal, and models in which a strong cardinal exists with no supercompact cardinal. These are not counterexamples inside ZFC to the implications just proved; they are separation results showing that the reverse implications require additional strength. Thus the comparison theorem is not a classification theorem and does not prove reversals from the axioms alone. Since implication alone does not describe separation in consistency strength, the next remark records what is known in the broader background of the course. [remark: Consistency Strength] The implications above are implications of large cardinal axioms. In the background consistency-strength ordering studied in this course, supercompactness is far stronger than strongness, and strongness is far stronger than measurability. Proving non-reversibility requires inner model theory or forcing methods beyond the present chapter. [/remark] The comparison also clarifies why the course treats supercompactness as a robustness principle. A supercompact cardinal can generate embeddings with closure requirements tailored to many different mathematical contexts. ## Menas Functions And Laver Functions Once embeddings become central, it is natural to ask whether we can predict or guide their behaviour. Menas and Laver functions are combinatorial devices that anticipate targets of embeddings and express a striking robustness of supercompactness. [definition: Menas Function] Let $\kappa$ be supercompact. A function $f:\kappa\to\kappa$ is a Menas function for $\kappa$ if for every ordinal $\lambda\geq\kappa$ there is a $\lambda$-supercompactness embedding $j:V\to M$ such that \begin{align*} j(f)(\kappa)>\lambda. \end{align*} [/definition] Thus the allowed embeddings are precisely the supercompactness embeddings from the previous sections. This motivates the standard Menas-function method: for a supercompact cardinal, one can often build a single small function on $\kappa$ whose value at $\kappa$ becomes arbitrarily large after choosing a sufficiently closed supercompactness embedding. The technical proofs of such lemmas use derived measures and comparison of ultrapowers, so in these notes the method is used as motivation rather than as a separate theorem card. The method needs supercompactness because it repeatedly chooses embeddings with arbitrarily high closure and then derives coherent measures from their seeds. A concrete boundary case is a model in which $\kappa$ is measurable but not $\kappa^+$-supercompact. Such a model may still have normal measures on $\kappa$, but it has no fine normal $\kappa$-complete measure on $P_\kappa(\kappa^+)$; by the fine normal measure theorem, that missing measure is exactly the obstruction to running the seed argument at $\lambda=\kappa^+$. In that situation no proof using only measurability can produce a function whose value is forced above every prescribed $\lambda$ by $\lambda$-supercompactness embeddings, because the relevant embeddings do not exist. Strongness gives a different precise limitation. If $\kappa$ is strong but not supercompact, then for some $\lambda$ there is no $\lambda$-supercompactness embedding, even though rank-initial embeddings may exist. The Menas property quantifies over $\lambda$-supercompactness embeddings, so it cannot be recovered from strongness alone at such a failed $\lambda$. This is stronger than the qualitative statement that sequence closure is useful: the exact missing hypothesis is the fine normal measure, or equivalently the closed ultrapower target, at the length under discussion. The theorem also predicts only ordinal height, not arbitrary sets or names. A function with the Menas property may tell us that an embedding can be chosen so that $j(f)(\kappa)$ is very large, but it does not let us prescribe the actual value of $j(f)(\kappa)$. The next question asks for a sharper anticipation principle: can one fixed function name arbitrary target objects, not only large ordinals, after passing to a suitable embedding? To state the sharper anticipation principle, we need a size bound on the target object. Here $H_{\lambda^+}$ denotes the sets whose transitive closures have cardinality below $\lambda^+$; equivalently, it is the hereditary-size bound appropriate to targets of size at most $\lambda$. This bound keeps the target object inside the range naturally controlled by a $\lambda$-supercompactness embedding. [definition: Laver Function] Let $\kappa$ be supercompact. A function $\ell:\kappa\to V_\kappa$ is a Laver function for $\kappa$ if for every set $a$ and every ordinal $\lambda$ with $a\in H_{\lambda^+}$, there is a $\lambda$-supercompactness embedding $j:V\to M$ such that \begin{align*} j(\ell)(\kappa)=a. \end{align*} [/definition] A Laver function is a preparation device: before choosing the embedding, the function $\ell$ is fixed once and for all, yet after applying a suitable embedding it can name any prescribed target object. The condition $a\in H_{\lambda^+}$ means that $\lambda$ is chosen large enough to cover the hereditary size of the desired target, and the chosen embedding target contains $a$ because $j(\ell)(\kappa)$ is an element of that target. The next theorem gives a rank-bounded version of this anticipation principle; taking a cardinal $\theta\geq\lambda$ with $a\in V_\theta$ recovers the hereditary-size formulation. [quotetheorem:7425] [proofunderconstruction:7425] The theorem is stronger than the Menas-function result because it anticipates actual sets rather than merely forcing an ordinal value to be large. The rank parameter is not cosmetic: as the target object becomes more complicated, the embedding must be chosen at a correspondingly large supercompactness length. If the target is originally specified by hereditary size, choose a cardinal $\theta$ above both the relevant $\lambda$ and the rank of the transitive closure of $a$; then $\theta$-supercompactness implies the weaker $\lambda$-supercompactness requirements. The conclusion is also existential in the right way. It says that for each target there is a suitable embedding, not that every embedding sends $\ell$ to that target. This is why Laver functions become useful as bookkeeping devices. They let later constructions choose an object, such as a forcing name, and then find a supercompactness embedding whose image of the fixed function names that object at $\kappa$. In its usual forcing application, Laver preparation uses this anticipation to make the supercompactness of $\kappa$ robust under broad classes of $<\kappa$-directed closed forcing, while arbitrary forcing such as a collapse can still destroy the large-cardinal structure. [example: Using A Laver Function As A Bookkeeping Device] Suppose a later forcing construction has a name $\dot{Q}$ for a $<\kappa$-directed closed forcing. Choose an ordinal $\lambda\geq\kappa$ large enough that $\dot{Q}\in H_{\lambda^+}$. If $\ell:\kappa\to V_\kappa$ is a Laver function for $\kappa$, then the definition of Laver function applies to the set $a=\dot{Q}$ and this ordinal $\lambda$: there is a $\lambda$-supercompactness embedding $j:V\to M$ such that \begin{align*} j(\ell)(\kappa)=\dot{Q}. \end{align*} The value is well-typed because $\operatorname{crit}(j)=\kappa$ and $j(\kappa)>\lambda\geq\kappa$, so $\kappa<j(\kappa)$ and therefore $\kappa$ is in the domain of $j(\ell)$. Since $j(\ell)$ is an element of $M$ and $M$ is transitive, the equality above places the forcing name $\dot{Q}$ inside the target model as the object selected at stage $\kappa$. Thus the Laver function turns the later forcing name into something the target model has already booked at its $\kappa$th stage, which is exactly the bookkeeping role used in the preparation argument. [/example] The chapter therefore ends with a shift in perspective. Strong cardinals show that elementary embeddings can make the target contain large rank initial segments; supercompact cardinals show that the target can be closed under long sequences; Laver functions show that these embeddings can be chosen with remarkable predictive power. Chapter 6 keeps the embedding language but distributes the critical points below a single cardinal, leading to Woodin cardinals. Strong and supercompact cardinals show how far the embedding viewpoint can be pushed by controlling the target model. Woodin cardinals keep embeddings at the center, but reorganize them around many critical points below one cardinal in order to control definability and forcing. # 6. Woodin Cardinals Woodin cardinals enter the course at the point where the elementary-embedding method stops being only a hierarchy of stronger closure properties and starts controlling definability in the universe. Earlier chapters treated inaccessible, measurable, and strong cardinals as reflection principles concentrated at a single critical point. This chapter studies a cardinal $\delta$ whose strength is spread throughout all of $V_\delta$: every subset of $V_\delta$ is reflected by some smaller critical point, and this uniform pattern is what makes Woodin cardinals central in determinacy and generic absoluteness. ## Projective Determinacy and Generic Absoluteness What kind of large cardinal hypothesis can explain why complicated definable sets of reals should have regularity properties, and why forcing should fail to change certain first-order truths? Descriptive set theory suggests that projective determinacy is not a local compactness phenomenon; it requires iterable elementary embeddings that can absorb arbitrary information coded below a cardinal. The Woodin cardinal is designed to provide exactly this repeated supply of strongness relative to every parameter below $\delta$. [motivation] ### Determinacy as Stability of Definable Games A projective set of reals is built from open sets by alternating projection and complement finitely many times. Games whose payoff sets are projective therefore encode finite but complicated patterns of quantification over reals. Projective determinacy asserts that every such game is determined, meaning one of the two players has a winning strategy. Large cardinals enter because a winning strategy is a highly absolute object: if it exists in one sufficiently correct universe, it should still be detected in related universes. Measurable cardinals give elementary embeddings, but a single measure concentrates on subsets of one cardinal and does not provide enough flexibility for arbitrary projective parameters. Woodin cardinals supply many critical points below $\delta$, each adapted to the particular set of information currently being reflected. ### Generic Absoluteness as Stability under Forcing Forcing changes the universe by adding sets, often reals. A generic absoluteness theorem says that certain statements, usually about $H_{\omega_1}$ or about projective sets of reals, have the same truth value before and after forcing. Such results need embeddings that can be lifted or compared across forcing extensions. The guiding point is that a Woodin cardinal behaves like a reservoir of strongness. Given a set $A \subset V_\delta$ coding a forcing construction, a payoff set, or a theory fragment, some $\kappa < \delta$ reflects $A$ far enough to preserve the relevant structure. [/motivation] The motivational slogans should be kept separate from the formal definition. The definition is internal to $V_\delta$ and speaks about strongness relative to an arbitrary subset $A \subset V_\delta$; determinacy and absoluteness are major consequences rather than part of the definition. [quotetheorem:7426] This theorem is stated here as a guide to the role of the concept. The proof belongs to descriptive set theory and inner model theory: it uses iteration trees, scales, and comparison arguments rather than only the elementary-embedding calculations developed in this course. Both hypotheses matter. With too few Woodin cardinals the known Martin-Steel argument does not reach the next projective level, and ZFC alone cannot prove even the first substantial projective determinacy consequences. The measurable cardinal above the finite Woodin sequence supplies the additional closure used to run the comparison and scale construction; omitting it is not a harmless cosmetic change in this quoted theorem, although sharper variants exist in more advanced formulations. [remark: Why the Statement Is Deliberately Schematic] The exact number of Woodin cardinals needed depends on the chosen indexing convention for projective pointclasses and on whether the statement is formulated with a measurable cardinal above the Woodin cardinals. The course uses this result as motivation for the definition and not as a theorem to prove in full. [/remark] This motivation points to a new kind of large cardinal axiom. Instead of asking for one embedding witnessing strength of $\delta$ itself, we ask that every parameter below $\delta$ be captured by many smaller embeddings. ## Strongness Relative to a Predicate How can an elementary embedding preserve not only the rank-initial segment $V_\lambda$, but also a chosen set $A$ of information living below $\delta$? Ordinary $\lambda$-strongness says that $V_\lambda \subset M$ for an embedding $j: V \to M$ with critical point $\kappa$. For Woodin cardinals we need a version where the embedding also respects a predicate $A$ up to $\lambda$. [definition: Critical Point] Let $j: V \to M$ be an elementary embedding, where $M$ is a transitive class. The critical point of $j$ is \begin{align*} \operatorname{crit}(j) := \min\{\alpha \in \operatorname{Ord} : j(\alpha) \ne \alpha\}. \end{align*} [/definition] The critical point measures where the embedding first moves the universe. Strongness below a Woodin cardinal is witnessed by embeddings whose critical points vary below $\delta$. [definition: A-Strongness] Let $A \subset V$ be a class or set, let $\kappa$ be a cardinal, and let $\lambda > \kappa$ be an ordinal. The cardinal $\kappa$ is $\lambda$-$A$-strong if there is an elementary embedding $j: V \to M$ such that: 1. $M$ is a transitive class; 2. $\operatorname{crit}(j)=\kappa$; 3. $V_\lambda \subset M$; 4. $j(A) \cap V_\lambda = A \cap V_\lambda$. [/definition] For a proper class predicate $A$, the notation $j(A)$ means the image predicate interpreted externally: $x \in j(A)$ iff, in the target structure, $x$ satisfies the formula or class parameter obtained by applying $j$ to the definition of $A$. Equivalently, one may regard the embedding as acting between structures in the expanded language with a unary predicate for $A$. The final condition says that $j$ preserves this predicate up to rank $\lambda$. When $A$ codes a structure, a forcing notion, or a sequence of parameters below $\delta$, the embedding is required to be strong without disturbing that code below the target rank. [example: Finding an A-Strong Critical Point] Fix $A \subset V_\delta$, and suppose $\delta$ satisfies the Woodin-cardinal condition. The order of choices is \begin{align*} A \subset V_\delta \quad\longmapsto\quad \kappa<\delta \quad\longmapsto\quad \text{for each }\lambda\text{ with }\kappa<\lambda<\delta,\text{ an embedding }j_\lambda:V\to M_\lambda . \end{align*} For each such $\lambda$, the embedding $j_\lambda$ must satisfy \begin{align*} \operatorname{crit}(j_\lambda)=\kappa,\qquad V_\lambda\subset M_\lambda,\qquad j_\lambda(A)\cap V_\lambda=A\cap V_\lambda . \end{align*} This explains why the target height cannot be chosen first without care. If $\lambda<\kappa$, then $\kappa$ cannot be $\lambda$-$A$-strong, because $\lambda$-$A$-strongness requires $\lambda>\kappa$. If $\kappa<\lambda<\delta$, then the same critical point $\kappa$ chosen for $A$ must work at that height, although the particular embedding may depend on $\lambda$. Thus $A$ is fixed first, one critical point $\kappa$ is then selected for that $A$, and every higher target rank $V_\lambda$ below $V_\delta$ must be reached by an embedding preserving $A$ through $V_\lambda$. [/example] The example shows why this is stronger than merely having many strong cardinals below $\delta$. Suppose, for instance, that below $\delta$ there are many cardinals strong to various heights, but for each such $\kappa$ a predicate $A \subset V_\delta$ codes the first stage at which the relevant embedding fails to preserve a chosen sequence. Then the supply of ordinary strong cardinals does not give a single critical point adapted to this $A$. Woodinness rules out exactly this diagonal obstruction by requiring the witness to be selected after $A$ is fixed. [definition: Woodin Cardinal via A-Strongness] A cardinal $\delta$ is Woodin if for every set $A \subset V_\delta$ there is a cardinal $\kappa < \delta$ such that for every ordinal $\lambda$ with $\kappa < \lambda < \delta$, $\kappa$ is $\lambda$-$A$-strong. [/definition] The same $\kappa$ must work for all higher target heights $\lambda < \delta$ once $A$ has been fixed. This uniformity is the central point of the definition: $\delta$ is not just a limit of strong behaviour, but a limit of strong behaviour tailored to every possible predicate on $V_\delta$. [remark: Quantifier Order] The definition has the form \begin{align*} \forall A \subset V_\delta\; \exists \kappa < \delta\; \forall \lambda\,(\kappa < \lambda < \delta \implies \exists j\; \Phi(A,\kappa,\lambda,j)). \end{align*} Changing the order of $\exists \kappa$ and $\forall \lambda$ weakens the statement. The critical point is chosen once for the predicate $A$, while the embedding may depend on the target height $\lambda$. [/remark] The next task is to connect this definition with the more common formulations in terms of functions $f: \delta \to \delta$ and elementary embeddings with $j(f)(\kappa) < \delta$ or $j(f)(\kappa) < j(\kappa)$. The reason functions enter is that a single $f$ can be used as an obstruction record: at each possible critical point it stores the first height where a desired preservation property fails. Closure points of $f$ are then candidates at which this record cannot move earlier information above the critical point. The embedding condition turns this bookkeeping device into a contradiction, because $j(f)(\kappa)$ is supposed to describe a first failure while the embedding itself witnesses preservation up to that height. ## Elementary-Embedding Characterisations Which formulation is most convenient in practice: preservation of every predicate $A \subset V_\delta$, or closure properties of embeddings associated to functions on $\delta$? The predicate formulation explains the meaning of Woodinness, while the function formulation is often easier to use in proofs. The equivalence is one of the basic technical facts about Woodin cardinals. [definition: Closure Point of a Function] Let $f: \delta \to \delta$ be a function and let $\kappa < \delta$. The ordinal $\kappa$ is a closure point of $f$ if \begin{align*} f``\kappa \subset \kappa. \end{align*} [/definition] The closure-point definition isolates the initial-segment control needed in the function formulation. It also creates the next problem: if a function can record how each candidate critical point fails to be strong for a predicate, then a closure point for that function should force the failure record to collapse. The equivalence theorem makes this coding argument precise. [quotetheorem:7427] [citeproof:7427] This theorem is the main working bridge for the chapter, but each condition has a job. The closure-point requirement $f``\kappa \subset \kappa$ prevents the obstruction function from smuggling information from below $\kappa$ to or above the critical point, which would make the diagonal argument fail. A concrete failure occurs if $f(\alpha)=\alpha+1$: then no nonzero $\kappa$ can be a closure point, so an embedding with critical point $\kappa$ would not tell us that the initial segment below $\kappa$ is closed under the obstruction record. The bound $j(f)(\kappa)<\delta$ is equally important: without it the alleged first failure could sit beyond the part of the universe controlled by Woodinness below $\delta$. The theorem therefore does not say that one embedding handles every predicate at once; it says that every attempted obstruction function has some critical point where the obstruction collapses. This is why the function form is best for deriving closure and reflection consequences. [example: A Function Measuring Failure of Strongness] Let $A \subset V_\delta$. Define $f:\delta\to\delta$ by \begin{align*} f(\gamma)=\min\{\lambda<\delta:\gamma<\lambda\text{ and }\gamma\text{ is not }\lambda\text{-}A\text{-strong}\} \end{align*} when this set is nonempty, and by $f(\gamma)=0$ when it is empty. Thus $f(\gamma)=0$ means that no failure of $A$-strongness for $\gamma$ occurs below $\delta$, while $f(\gamma)=\theta>0$ means \begin{align*} \gamma<\theta<\delta \end{align*} and $\theta$ is the first target height below $\delta$ at which $\gamma$ fails to be $\theta$-$A$-strong. Now suppose the function characterisation is applied to this obstruction code and gives an elementary embedding $j:V\to M$ with $\operatorname{crit}(j)=\kappa$, $V_{j(f)(\kappa)}\subset M$, and $j(f)(\kappa)<\delta$. Put \begin{align*} \theta=j(f)(\kappa). \end{align*} Inside $M$, the value $\theta$ is computed from $j(f)$ by the same least-failure rule, so if $\theta>0$, then $M$ reads $\theta$ as the least ordinal such that $\kappa$ is not $\theta$-$j(A)$-strong. But the embedding $j$ itself has critical point $\kappa$, and the condition $V_\theta\subset M$ gives the required rank closure at exactly that height. In the diagonal application, the code is chosen so that the relevant predicate is preserved below $\theta$, namely \begin{align*} j(A)\cap V_\theta=A\cap V_\theta. \end{align*} Therefore the same embedding data witnesses the strongness that $\theta$ was supposed to be the first failure of. Hence $\theta$ cannot be a positive least failure. The only consistent outcome is that the obstruction has collapsed at $\kappa$: there is no first failure below $\delta$ for the predicate being coded. [/example] The function formulation also clarifies how Woodinness differs from measurability. A measurable cardinal supplies an elementary embedding from an ultrapower, but it does not assert that below a larger cardinal there are critical points strong for every predicate. ## Woodin Cardinals as Limits of Strongness Patterns In what sense is a Woodin cardinal a limit of large cardinals below it? The answer is subtler than saying that there are many strong cardinals below $\delta$. Unpacking the definition, for each predicate $A \subset V_\delta$, Woodinness produces a critical point $\kappa$ that is strong to every height below $\delta$ while preserving $A$; as $A$ varies, this creates a dense pattern of relative strongness below $\delta$. This result is often the first way to use Woodinness in calculations, but its content is not just that some strong cardinal exists below $\delta$. If the hypothesis were weakened to "for each $\lambda<\delta$ there is some $\kappa$ strong to $\lambda$," the critical point could change with $\lambda$, and no single embedding pattern would preserve a fixed predicate $A$ through all higher target ranks. A specific obstruction is obtained by letting $A$ code, for each candidate $\kappa$, the least height at which the chosen ordinary strongness witness for $\kappa$ fails to preserve a fixed list of parameters; ordinary strongness supplies embeddings, but not embeddings respecting this diagonal predicate. The predicate-relative condition is also necessary: an embedding that is $\lambda$-strong for the empty predicate may move the coded set $A\cap V_\lambda$, so it cannot be used in arguments where $A$ records a forcing notion, a strategy, or an iteration tree. The theorem has a boundary as well: it gives embeddings depending on $\lambda$, not one master embedding covering every $V_\lambda$ at once. [example: Comparing a Measurable Cardinal with a Woodin Cardinal] A measurable cardinal $\mu$ gives one elementary embedding \begin{align*} j:V\to M \end{align*} obtained from a normal measure on $\mu$, with \begin{align*} \operatorname{crit}(j)=\mu. \end{align*} This means that $j(\alpha)=\alpha$ for every ordinal $\alpha<\mu$, while $j(\mu)\ne\mu$. The critical point is therefore fixed in advance by the measure: it is $\mu$. Woodinness has a different quantifier pattern. For a Woodin cardinal $\delta$, after a predicate $A\subset V_\delta$ is chosen, one must find a cardinal \begin{align*} \kappa<\delta \end{align*} such that for every ordinal $\lambda$ satisfying \begin{align*} \kappa<\lambda<\delta, \end{align*} there is an elementary embedding \begin{align*} j_\lambda:V\to M_\lambda \end{align*} with \begin{align*} \operatorname{crit}(j_\lambda)=\kappa,\qquad V_\lambda\subset M_\lambda,\qquad j_\lambda(A)\cap V_\lambda=A\cap V_\lambda. \end{align*} Thus measurability supplies one critical point and one measure-derived embedding pattern at $\mu$, while Woodinness supplies, for each predicate $A\subset V_\delta$, a critical point below $\delta$ whose embeddings preserve that particular predicate through every higher rank $V_\lambda$ below $V_\delta$. This is why a Woodin cardinal is a distributed predicate-reflection principle rather than a single-ultrafilter principle. [/example] The preceding comparison prevents a common misreading. Woodinness is not simply measurability plus closure; it is a uniform scheme producing strongness witnesses below the cardinal for all possible predicates on its rank-initial segment. [quotetheorem:7428] [citeproof:7428] Inaccessibility shows that a Woodin cardinal has the usual closure expected of a large cardinal, but it is only a coarse shadow of Woodinness. An inaccessible cardinal may have no predicate-relative strongness pattern below it; for instance, in a model of $V=L$ with an inaccessible cardinal, Scott's theorem rules out measurability and hence rules out the embedding strength needed for Woodinness. Regularity and the strong-limit property say only that cofinal sequences and power sets below $\delta$ remain bounded below $\delta$. The stronger takeaway is that Woodinness makes $\delta$ a limit of predicate-relative strongness patterns: for each fixed $A\subset V_\delta$ and each lower bound $\alpha<\delta$, there are critical points $\kappa$ with $\alpha<\kappa<\delta$ that witness $A$-strongness through arbitrarily high ranks below $\delta$. This explains the phrase "limit of strongness patterns," and its hypotheses are doing real work. If we only knew that one witness existed for each predicate, all such witnesses could in principle lie below a fixed bound $\alpha$. The closure-point trick in the usual proof rules this out by building the lower bound directly into the obstruction function. A concrete boundary case is a model with a single strong cardinal $\kappa$ below an inaccessible $\delta$ and no comparable strongness witnesses above $\kappa$: it may give substantial reflection at $\kappa$, but it cannot make the witnesses unbounded in $\delta$. The conclusion is still predicate-relative: it does not produce a club of witnesses that works uniformly for every $A \subset V_\delta$ at once, but it does give an unbounded supply once the predicate has been fixed. ## Finite Woodin Cardinals and Projective Consequences How much of the Woodin-cardinal machinery is needed to obtain consequences about sets of reals? A single Woodin cardinal already gives a substantial generic absoluteness principle for low-complexity statements, while finite sequences of Woodin cardinals correspond to increasingly complex projective levels. The course treats this as a consistency-strength guide rather than developing the full descriptive set-theoretic proof. [definition: Finite Sequence of Woodin Cardinals] A finite sequence of Woodin cardinals is a sequence \begin{align*} \delta_1 < \delta_2 < \dots < \delta_n \end{align*} where each $\delta_i$ is Woodin. [/definition] The ordering matters because applications usually require iterating embeddings and comparing models from the bottom Woodin cardinal upward. A measurable cardinal above the finite sequence often supplies the final amount of closure needed for the standard projective determinacy theorem. [quotetheorem:7429] This theorem is quoted without proof in this course. Its proof uses the full machinery of iteration strategies and inner models for Woodin cardinals, which lies beyond the embedding calculations needed to understand the definition. The finite sequence is not decorative: the number of Woodin cardinals controls the projective level reached by the theorem, so removing one Woodin cardinal drops the known conclusion to a lower level. The measurable cardinal above the sequence supplies the extra closure used in the standard quoted form; without that closure the proof scheme stated here no longer applies. At the opposite extreme, having only inaccessible or measurable cardinals does not yield projective determinacy by the embedding calculations in these notes, since those cardinals do not provide the predicate-relative sequence of strongness witnesses needed for the scale construction. The regularity conclusions are the concrete payoff. Once the relevant projective games are determined, the associated pointclasses inherit Lebesgue measurability, the Baire property, and the perfect set property by the determinacy regularity theorems quoted earlier. For example, the theorem says that a finite supply of Woodin cardinals does not merely produce another embedding statement: it explains why definable sets of reals at a corresponding projective level behave as though they cannot encode arbitrary pathological counterexamples. The limitation is just as important. The theorem calibrates finite projective levels, not all sets of reals and not full determinacy for every payoff set. Increasing $n$ raises the definability complexity that can be controlled, while the proof methods require inner-model comparison and iteration strategies not developed here. That is why this section uses the result as a bridge: it connects the internal definition of Woodinness to visible consequences in descriptive set theory, then returns to the large-cardinal hierarchy rather than trying to reproduce the full Martin-Steel proof. [remark: Consistency Strength] The projective consequences should be read as relative consistency statements in the large-cardinal hierarchy. The point is not that ZFC proves projective determinacy, but that specific finite large-cardinal hypotheses explain why projective determinacy is consistent relative to those hypotheses. [/remark] The chapter therefore ends with two takeaways. Formally, a Woodin cardinal is a cardinal $\delta$ such that every predicate $A \subset V_\delta$ has an unbounded supply of strongness witnesses below $\delta$. Conceptually, this distributed strongness is what lets Woodin cardinals govern the interaction between large cardinals, definable sets of reals, and forcing absoluteness. Chapter 7 turns to the obstruction side of the story: why the constructible universe $L$ cannot by itself contain the witnesses needed for this embedding strength. Woodin cardinals connect embedding strength with the structure of definable sets of reals. The next chapter steps back to examine the canonical inner model $L$, whose rigidity and minimality explain both why it is powerful and why it cannot accommodate the large-cardinal witnesses developed so far. # 7. The Constructible Universe and Obstacles to Large Cardinals The previous chapters developed large cardinals through reflection, filters, ultrapowers, and elementary embeddings. This chapter pauses to ask why these principles cannot be studied only inside the constructible universe $L$, even though $L$ is the most canonical inner model of ZFC. The tension is that $L$ has a definable fine structure and a global well-order, while large cardinals tend to produce embeddings, indiscernibles, and covering failures that cannot be captured by naive constructibility. ## The Role of $L$ in Testing Large Cardinal Strength The guiding question is: what does the constructible universe preserve from $V$, and what does it destroy? For ordinary set-theoretic questions, $L$ is a useful minimal universe; for large cardinal theory, it becomes a testing ground where many proposed large-cardinal features disappear. [definition: Constructible Hierarchy] The constructible hierarchy is the class sequence $(L_\alpha)_{\alpha \in \operatorname{Ord}}$ defined by transfinite recursion as follows: \begin{align*} L_0 = \varnothing. \end{align*} \begin{align*} L_{\alpha+1} = \operatorname{Def}(L_\alpha). \end{align*} \begin{align*} L_\lambda = \bigcup_{\alpha < \lambda} L_\alpha \quad \text{for limit } \lambda. \end{align*} Here $\operatorname{Def}(L_\alpha)$ is the collection of subsets of $L_\alpha$ first-order definable over $(L_\alpha, \in)$ with parameters from $L_\alpha$. The constructible universe is \begin{align*} L = \bigcup_{\alpha \in \operatorname{Ord}} L_\alpha. \end{align*} [/definition] The hierarchy explains why $L$ is canonical, but it also exposes a problem: large cardinal witnesses are often extra objects, such as measures or extender fragments, rather than ordinals alone. This motivates a controlled way of adding a predicate while keeping the constructible method. [definition: Relativized Constructibility] For a class $A$, the relativized constructible hierarchy $(L_\alpha[A])_{\alpha \in \operatorname{Ord}}$ is defined by transfinite recursion as follows: \begin{align*} L_0[A] = \varnothing. \end{align*} \begin{align*} L_{\alpha+1}[A] = \operatorname{Def}(L_\alpha[A], \in, A \cap L_\alpha[A]). \end{align*} \begin{align*} L_\lambda[A] = \bigcup_{\alpha < \lambda} L_\alpha[A] \quad \text{for limit } \lambda. \end{align*} The relativized constructible universe is \begin{align*} L[A] = \bigcup_{\alpha \in \operatorname{Ord}} L_\alpha[A]. \end{align*} [/definition] Relativized constructibility is the first hint that $L$ alone is too small. Inner model theory often tries to add carefully chosen predicates or extender sequences, not arbitrary sets, so that the resulting model remains canonical while still recording large-cardinal information. [example: Constructibility Versus Rank Initial Segments] Let $\kappa$ be inaccessible in $V$. We compute why $V_\kappa$ behaves like a full universe below $\kappa$, while $L_\kappa$ has the same ordinal height but only the subsets produced by the constructible recursion. Since $\kappa$ is a limit ordinal, if $x,y\in V_\kappa$, then there are $\alpha,\beta<\kappa$ with $x\in V_\alpha$ and $y\in V_\beta$; putting $\gamma=\max\{\alpha,\beta\}+2<\kappa$, we have $\{x,y\}\in V_\gamma$, so Pairing holds in $V_\kappa$. If $x\in V_\alpha$ for some $\alpha<\kappa$, then every member of $\bigcup x$ is a member of a member of $x$, hence belongs to some rank below $\alpha$; therefore $\bigcup x\subseteq V_\alpha$ and $\bigcup x\in V_{\alpha+1}\subseteq V_\kappa$. Infinity holds because $\omega\in V_{\omega+1}\subseteq V_\kappa$, and Extensionality and Foundation are inherited from the ambient transitive universe $V$. For Separation, let $a\in V_\kappa$ and let $b=\{x\in a:\varphi(x,\vec p)\}$, where the parameters $\vec p$ belong to $V_\kappa$. Since $b\subseteq a$, the rank of $b$ is at most the rank of $a$ plus $1$; choosing $\alpha<\kappa$ with $a\in V_\alpha$, we get $b\in V_{\alpha+1}\subseteq V_\kappa$. For Power Set, choose $\alpha<\kappa$ with $x\in V_\alpha$. If $y\subseteq x$, then every element of $y$ is an element of $x$, so $y$ has rank no larger than the rank of $x$; hence $y\in V_\alpha$. Thus \begin{align*} \mathcal P(x)\subseteq V_\alpha. \end{align*} Since $\mathcal P(x)$ is a set of elements of $V_\alpha$, it belongs to $V_{\alpha+1}$, and because $\alpha+1<\kappa$ we have $\mathcal P(x)\in V_\kappa$. Replacement uses regularity of $\kappa$. Let $a\in V_\kappa$, and let $F$ be a definable class function on $a$ with values in $V_\kappa$. Since $\kappa$ is strongly inaccessible, every rank $V_\alpha$ for $\alpha<\kappa$ has cardinality less than $\kappa$, so $|a|<\kappa$. For each $x\in a$, choose $\alpha_x<\kappa$ such that $F(x)\in V_{\alpha_x}$. Regularity gives \begin{align*} \beta=\sup\{\alpha_x+1:x\in a\}<\kappa. \end{align*} Then every value $F(x)$ lies in $V_\beta$, so \begin{align*} F``a\subseteq V_\beta. \end{align*} Therefore $F``a\in V_{\beta+1}\subseteq V_\kappa$, which is exactly the Replacement witness inside $V_\kappa$. The constructible segment $L_\kappa$ has ordinal height $\kappa$, but its subsets are filtered through the stages of the constructible hierarchy. If $\alpha<\kappa$, then transitivity of $L_\kappa$ gives \begin{align*} \mathcal P(\alpha)^{L_\kappa}=\{A\in L_\kappa:A\subseteq \alpha\}. \end{align*} The right-hand side is exactly \begin{align*} \mathcal P(\alpha)\cap L_\kappa. \end{align*} Since \begin{align*} L_\kappa=\bigcup_{\beta<\kappa}L_\beta, \end{align*} a subset $A\subseteq\alpha$ belongs to $L_\kappa$ precisely when $A\in L_\beta$ for some $\beta<\kappa$. Thus $V_\kappa$ contains every subset of each earlier set that exists in $V$, while $L_\kappa$ contains only those subsets constructed before stage $\kappa$. In particular, if a normal measure $U$ on $\kappa$ exists in $V$ but $U\notin L$, then the ordinal $\kappa$ is present in the constructible hierarchy, while the measure witnessing measurability is not constructible. The comparison shows that matching ordinal height does not determine matching power sets, matching witnesses, or matching large-cardinal structure. [/example] The comparison between $V_\kappa$ and $L_\kappa$ raises a stability question: when different transitive universes compute constructibility, do they obtain the same initial levels? The obstruction is that constructibility is defined by an internal recursion using definability over earlier stages, so a nontransitive or ordinal-deficient model could compute the wrong subsets or fail to run the recursion far enough. The theorem below isolates the hypotheses under which the initial levels of $L$ are canonical across ambient models. [quotetheorem:7430] [citeproof:7430] The hypotheses are doing real work. Transitivity prevents $M$ from having nonstandard membership relations that would make formulas over an apparent level $L_\beta$ define the wrong subsets, and containing the relevant ordinals prevents the internal recursion from stopping too soon or omitting stages. For a concrete boundary case, take a countable elementary submodel $N \prec H_\theta$ containing a countable ordinal $\alpha$ and collapse it to a transitive model $\bar N$. The uncollapsed $N$ is not transitive: an element of a set in $N$ need not itself belong to $N$, so definability over an internal copy of a level can miss parameters and members that exist externally. Conversely, if a transitive model has ordinal height $\gamma$ and one asks it to compute a level beyond $\gamma$, the recursion has no internal stage corresponding to the missing ordinals. The theorem therefore does not say that $M$ and $V$ have the same sets, nor that $M$ computes all of $L$ correctly beyond its ordinal height; it says only that the canonical construction is absolute on the levels whose ingredients are genuinely present. This precise stability is what later allows covering and sharp arguments to compare $L$ across different ambient universes. ## Covering and the Failure of Naive Measurability in $L$ Suppose $V$ has no sharp for $L$. How close must $L$ then be to $V$? Jensen's covering phenomenon answers this by saying that uncountable sets of ordinals in $V$ are nearly covered by constructible sets of the same size. [definition: Covering Property for $L$] The covering property for $L$ says that for every uncountable set $X \subseteq \operatorname{Ord}$, there is a set $Y \in L$ such that \begin{align*} X \subseteq Y \quad \text{and} \quad |Y| = |X|. \end{align*} [/definition] Covering means that $L$ has enough sets of ordinals to approximate $V$ at uncountable cardinalities. Jensen's fine-structural covering theorem supplies the dividing line used here: if $0^\sharp$ does not exist, then every uncountable set of ordinals is contained in a constructible set of the same cardinality. This is a deep external input from fine structure, not a theorem proved from the earlier material in this course. The hypothesis that $0^\sharp$ does not exist is essential, not cosmetic. Here is the standard failure mode in the presence of Silver indiscernibles. Let $I=\{i_\xi:\xi\in\operatorname{Ord}\}$ be the increasing class of Silver indiscernibles. Choose an index $\beta$ so large that $i_\beta$ has external cardinality greater than $\aleph_1^V$, and set \begin{align*} \lambda=\sup_{\xi<\omega_1^V} i_{\beta+\xi}, \qquad X=\{i_{\beta+\xi}:\xi<\omega_1^V\}. \end{align*} Then $X$ is an uncountable set of ordinals of size $\aleph_1^V$ and is cofinal in $\lambda$. The ordinal $\lambda$ is a [limit point](/page/Limit%20Point) of the Silver indiscernibles which is singular in $V$ with cofinality $\omega_1^V$, while the corresponding constructible cardinal computation treats the interval below $\lambda$ as regular: in $L$, every unbounded subset of $\lambda$ has constructible cardinality $\lambda$. If $Y\in L$ contains $X$, then $Y\cap\lambda$ is unbounded in $\lambda$. Applying this regularity statement inside $L$ to the constructible set $Y\cap\lambda$, there is in $L$ an injection from $\lambda$ into $Y\cap\lambda$, and the same injection exists externally. Hence \begin{align*} |Y|^V \ge |\lambda|^V > \aleph_1^V = |X|^V. \end{align*} No constructible $Y$ can both cover this specific $X$ and have $|Y|=|X|$. This gives a concrete obstruction: externally, $X$ is small because it is an $\omega_1^V$-sequence of indiscernibles, while any constructible cover must be large enough to cover a whole unbounded constructible interval forced by the indiscernibility pattern. Thus the lemma should be read as a dichotomy: either $L$ approximates uncountable sets of ordinals very tightly, or there is sharp-like structure witnessing that $L$ is missing coherent elementary information. Covering still does not assert $V=L$, and it says nothing directly about reals or about arbitrary witnesses such as measures; it is a controlled approximation theorem for uncountable sets of ordinals. [remark: Covering Is Not Equality] Covering is a size-preserving approximation property, not the assertion $V=L$. Forcing extensions can add new reals while still leaving certain uncountable covering patterns intact. The lemma is therefore a structural comparison theorem rather than a definability theorem. [/remark] Measurability now becomes delicate. A measure on $\kappa$ in $V$ need not belong to $L$, and the ordinal $\kappa$ alone does not encode the ultrafilter needed to form the ultrapower. The obstruction is therefore a witness problem: an inner model can contain the candidate cardinal while omitting the measure that makes it measurable. The theorem below states this failure explicitly. [quotetheorem:7431] [citeproof:7431] This theorem is a warning about witnesses rather than about the ordinal $\kappa$ alone. The model $V=L[U]$ is the guiding boundary case: the ambient universe has a normal measure because $U$ was deliberately added to the construction, while its submodel $L$ lacks that predicate and cannot reconstruct it from $\kappa$ alone. If a normal measure $U$ on $\kappa$ lies outside $L$, then $L$ has no reason to contain the ultrapower data needed to verify measurability internally, even though $\kappa$ is still an ordinal of $L$. The contrast is $L[U]$: when the predicate $U$ is built into the inner model in a coherent way, the model can form the corresponding ultrapower and may verify that $\kappa$ is measurable there. Thus measurability can be inherited by an inner model only when the relevant measure, or a structured code for it, is itself internal to that model. [example: A Measure Outside the Constructible Universe] Let $U$ be a normal measure on a measurable cardinal $\kappa$ in $V$, and restrict it to the constructible subsets of $\kappa$: \begin{align*} U^L=U\cap \mathcal P(\kappa)^L. \end{align*} Here $\kappa\in L$ because every ordinal is constructible, but $U$ is not determined by the ordinal $\kappa$ alone: it is a set whose elements are subsets of $\kappa$. To use this particular external measure inside $L$, the model $L$ would have to decide the predicate \begin{align*} A\mapsto A\in U \end{align*} for each $A\in\mathcal P(\kappa)^L$. Suppose, toward seeing what this would imply, that there are a formula $\theta(v,a)$ and a parameter $a\in L$ such that for every $A\in\mathcal P(\kappa)^L$, \begin{align*} A\in U \quad \text{if and only if} \quad L\models \theta(A,a). \end{align*} Since $\mathcal P(\kappa)^L$ is a set of $L$, Separation in $L$ forms the set \begin{align*} W=\{A\in\mathcal P(\kappa)^L:L\models \theta(A,a)\}. \end{align*} Using the displayed equivalence, the elements of $W$ are exactly the constructible subsets of $\kappa$ which belong to $U$, so \begin{align*} W=\{A\in\mathcal P(\kappa)^L:A\in U\}. \end{align*} By the definition of intersection, \begin{align*} \{A\in\mathcal P(\kappa)^L:A\in U\}=U\cap\mathcal P(\kappa)^L. \end{align*} Therefore \begin{align*} W=U^L. \end{align*} Because $W$ was formed by Separation inside $L$, this gives $U^L\in L$. Thus, if $U^L\notin L$, no parameter from $L$ can define the restriction of $U$ to constructible subsets of $\kappa$. This is exactly the missing information needed for the $U$-ultrapower: for constructible functions $f,g:\kappa\to L$, \begin{align*} f=_U g \quad \text{means} \quad \{\xi<\kappa:f(\xi)=g(\xi)\}\in U, \end{align*} and \begin{align*} [f]_U\in_U[g]_U \quad \text{means} \quad \{\xi<\kappa:f(\xi)\in g(\xi)\}\in U. \end{align*} The two displayed subsets of $\kappa$ are constructible whenever $f$ and $g$ are constructible, but deciding whether they belong to the external measure $U$ requires the missing predicate $U^L$. So $L$ may contain many subsets of $\kappa$ while still lacking the specific measure data needed to form the ultrapower by this particular external measure. [/example] The obstruction is not only that $L$ is small. The deeper issue is that the existence of measures and embeddings often produces canonical non-constructible information, and that information is organized by sharps. ## Silver's $0^\sharp$ and Indiscernibles for $L$ What is the first object that records a genuine elementary-embedding structure over $L$? Silver's $0^\sharp$ is the answer in the region before measurable-cardinal inner models: it is a real coding a complete theory of $L$ with a proper class of indiscernibles. [definition: Indiscernibles for $L$] A class $I \subseteq \operatorname{Ord}$ is a class of indiscernibles for $L$ if for every formula $\varphi(v_1, \dots, v_n)$ and any increasing sequences $\alpha_1 < \dots < \alpha_n$ and $\beta_1 < \dots < \beta_n$ from $I$, \begin{align*} L \models \varphi(\alpha_1, \dots, \alpha_n) \quad \iff \quad L \models \varphi(\beta_1, \dots, \beta_n). \end{align*} [/definition] Indiscernibles make the definable universe $L$ behave as if it has hidden symmetries, but a proper class of such ordinals is too large to handle as a single real parameter. The coding problem is to compress the elementary pattern of these indiscernibles into one object whose existence can be discussed inside ordinary set-theoretic arguments. Silver's $0^\sharp$ is that code: it records the complete theory that generates the indiscernibility structure over $L$. [definition: Silver's $0^\sharp$] Silver's $0^\sharp$, when it exists, is the set of Gödel numbers of formulas true in the canonical Ehrenfeucht-Mostowski theory for $L$ with constants $(c_i)_{i \in \omega}$ assigned to increasing order indiscernibles. [/definition] Equivalently, $0^\sharp$ is a real coding the complete theory whose well-founded Ehrenfeucht-Mostowski models generate a proper class of Silver indiscernibles for $L$. The definition packages a proper class phenomenon into a real. The next theorem is needed to turn that coded theory into the concrete structural statement used in inner model arguments. [quotetheorem:7432] [citeproof:7432] This theorem is the first serious sign that $L$ can be studied from outside by embedding-like data, but the assumption is substantial. ZFC alone neither proves nor refutes the existence of $0^\sharp$ relative to the usual consistency background; if it does not exist, there is no proper class of Silver indiscernibles for $L$. A concrete boundary is given by forcing over a model of $V=L$ to add a Cohen real: the extension has new nonconstructible sets, but it still has no $0^\sharp$, so mere failure of $V=L$ is not enough to create indiscernibles for $L$. The conclusion also does not place the indiscernibles inside $L$ as a class definable by $L$ itself. Rather, it gives an external class in $V$ whose order patterns generate elementary maps between constructible levels, and those maps are the ancestors of the extender sequences used in later inner model theory. [example: Skolem Hulls from Indiscernibles] Assume $I$ is a proper class of indiscernibles for $L$, and let $a=\{\alpha_0<\cdots<\alpha_{n-1}\}\subset I$ be finite. Fix external Skolem functions for $L$: for each formula $\varphi(y,\vec{x})$, let $F_\varphi(\vec{x})$ be the $<_L$-least $y$ such that $L\models\varphi(y,\vec{x})$, when such a $y$ exists. The Skolem hull generated by $a$ is \begin{align*} H_a=\{F_\varphi(\alpha_{i_0},\dots,\alpha_{i_{m-1}}):\varphi \text{ a formula, } i_0,\dots,i_{m-1}<n,\text{ and the value exists}\}. \end{align*} Take another increasing tuple $b=\{\beta_0<\cdots<\beta_{n-1}\}\subset I$ of the same order type. We define the correspondence on Skolem terms by \begin{align*} \pi(F_\varphi(\alpha_{i_0},\dots,\alpha_{i_{m-1}}))=F_\varphi(\beta_{i_0},\dots,\beta_{i_{m-1}}). \end{align*} To see that this does not depend on the chosen term representation, suppose \begin{align*} F_\varphi(\alpha_{i_0},\dots,\alpha_{i_{m-1}})=F_\psi(\alpha_{j_0},\dots,\alpha_{j_{\ell-1}}). \end{align*} This equality is equivalent to the first-order assertion over $L$ that there is a unique $<_L$-least $u$ satisfying $\varphi(u,\alpha_{i_0},\dots,\alpha_{i_{m-1}})$, a unique $<_L$-least $v$ satisfying $\psi(v,\alpha_{j_0},\dots,\alpha_{j_{\ell-1}})$, and $u=v$. Since the indices merely select entries from the increasing tuple $(\alpha_0,\dots,\alpha_{n-1})$, this is a first-order statement about that tuple from $I$. By indiscernibility, the same statement holds after replacing $(\alpha_0,\dots,\alpha_{n-1})$ by $(\beta_0,\dots,\beta_{n-1})$, so \begin{align*} F_\varphi(\beta_{i_0},\dots,\beta_{i_{m-1}})=F_\psi(\beta_{j_0},\dots,\beta_{j_{\ell-1}}). \end{align*} Thus $\pi$ is well-defined. It is onto $H_b$ because every element of $H_b$ has the form $F_\varphi(\beta_{i_0},\dots,\beta_{i_{m-1}})$, and this is the image under $\pi$ of $F_\varphi(\alpha_{i_0},\dots,\alpha_{i_{m-1}})$. It is one-to-one by applying the same equality argument in the reverse direction, from the tuple $b$ back to the tuple $a$. Finally, let $\theta(z_0,\dots,z_{r-1})$ be any formula, and choose elements of $H_a$ represented as \begin{align*} z_q=F_{\varphi_q}(\alpha_{i^q_0},\dots,\alpha_{i^q_{m_q-1}}) \quad \text{for } q<r. \end{align*} The assertion $L\models\theta(z_0,\dots,z_{r-1})$ can be rewritten as a first-order statement about $(\alpha_0,\dots,\alpha_{n-1})$ saying: the displayed Skolem definitions produce the objects $z_0,\dots,z_{r-1}$, and $\theta(z_0,\dots,z_{r-1})$ holds. By indiscernibility, the same statement holds for $(\beta_0,\dots,\beta_{n-1})$. Therefore \begin{align*} L\models\theta(z_0,\dots,z_{r-1}) \quad \text{if and only if} \quad L\models\theta(\pi(z_0),\dots,\pi(z_{r-1})). \end{align*} So $\pi:H_a\to H_b$ is an elementary correspondence, and the elementary structure of the Skolem hull depends only on the finite order pattern of the chosen indiscernibles, not on the particular ordinals used. [/example] The hull construction also explains why covering and sharps are opposite phenomena. The remaining question is how to move from a failed covering approximation to the sharp-like elementary structure just described. [quotetheorem:7433] [citeproof:7433] The boundary case is exactly the sharp threshold, but the theorem is deliberately narrower than "if $V\ne L$, then $0^\sharp$ exists." For a specific boundary example, start with a universe satisfying $V=L$ and force to add a Cohen real $c\subseteq\omega$. In the extension, $c$ is a new set of ordinals and $V\ne L$, but $0^\sharp$ still does not exist. This does not contradict the theorem, because the covering statement concerns uncountable sets of ordinals and the new real is countable; indeed $\omega\in L$ is a same-size constructible cover of $c$. By contrast, the Silver-indiscernible set $X=\{i_{\beta+\xi}:\xi<\omega_1^V\}$ described above is an uncountable set whose attempted constructible covers are forced by indiscernibility to be too large. The implication therefore detects a precise kind of uncountable approximation failure, not every possible way in which $L$ may fail to equal the ambient universe. ## Obstacles to Large Cardinals Inside $L$ Why not simply continue using $L$ as the canonical inner model for all large cardinal arguments? The answer is that $L$ is too rigid to carry the witnesses demanded by stronger large cardinals, and the first symptom is already visible at measurability. [quotetheorem:7434] [citeproof:7434] This obstruction identifies the next construction problem: build a canonical inner model that still resembles $L$ but carries the extra data needed for embeddings. The statement concerns the particular external measure data $U^L$, not every possible measure on $\kappa$. If $U_L \in L$ is a normal measure on $\kappa$ as computed by $L$, then $L$ can define the ultrapower relation on its own functions $f:\kappa \to L$, form the corresponding internal ultrapower class, and obtain the usual ultrapower embedding with critical point $\kappa$. If one works instead in $L[U]$ or in a coherent mouse whose extender sequence contains the relevant measure information, the same kind of construction becomes internal to that richer model. The point is therefore not that ultrapowers are incompatible with constructibility, but that bare $L$ has no predicate recording a witness that lies outside it. The term used for the fine-structural candidates is premouse. [definition: Premouse] A premouse is a transitive fine-structural structure of the form \begin{align*} M = (J_\alpha^{\vec{E}}, \in, \vec{E}) \end{align*} where $J_\alpha^{\vec{E}}$ is a Jensen-style level built relative to an extender sequence $\vec{E}$. The sequence $\vec{E}$ is required to be amenable to the structure and coherent with the initial segments of the construction. [/definition] At this stage, "fine-structural", "amenable", and "coherent" are placeholders for technical requirements developed later. Amenability means that initial pieces of the extender sequence are visible to the corresponding initial segments of $M$; coherence means that ultrapowers and initial segments agree about which extender data has already been used. The definition is schematic, but the point needed here is precise: a premouse is $L$ rebuilt with a controlled predicate for measure or extender information. Premice are not introduced for their own sake. They are the technical response to the failure of $L$: if measures and embeddings are real mathematical witnesses, a canonical inner model must contain structured codes for them. [remark: Why Extenders Replace Single Measures] A single normal measure is adequate for the first ultrapower of a measurable cardinal, but stronger large cardinals require coherent systems of measures indexed over many finite sequences and critical points. Extenders package these systems so that the inner model can remember enough of the elementary embedding to be iterable and comparable with other models. [/remark] The comparison with $L$ remains essential. Fine-structural models are judged by whether they preserve the virtues of $L$: definability, condensation, canonical construction, and comparison. They differ from $L$ by adding the minimum structured information needed to support large cardinals. [example: From $L$ to $L[U]$] Let $U$ be a normal measure on $\kappa$ in the ambient universe, and build the relativized hierarchy by \begin{align*} L_0[U]=\varnothing. \end{align*} \begin{align*} L_{\alpha+1}[U]=\operatorname{Def}(L_\alpha[U],\in,U\cap L_\alpha[U]). \end{align*} \begin{align*} L_\lambda[U]=\bigcup_{\alpha<\lambda}L_\alpha[U] \quad \text{for limit } \lambda. \end{align*} Thus \begin{align*} L[U]=\bigcup_{\alpha\in\operatorname{Ord}}L_\alpha[U]. \end{align*} The difference from bare $L$ is that each successor stage may define sets using the predicate recording the part of $U$ already visible at the previous stage. In particular, for constructible-from-$U$ subsets of $\kappa$, the relevant measure is \begin{align*} U^{L[U]}=U\cap \mathcal P(\kappa)^{L[U]}. \end{align*} If the construction is coherent, this restricted measure is not an external afterthought: it is coded by the predicate sequence of the model, so $L[U]$ can decide, for each $A\in\mathcal P(\kappa)^{L[U]}$, whether \begin{align*} A\in U^{L[U]}. \end{align*} This is exactly the information missing from $L$ when $U\notin L$. For $f,g:\kappa\to L[U]$ belonging to $L[U]$, the equality relation in the ultrapower is defined by \begin{align*} f=_{U^{L[U]}}g \quad \text{if and only if} \quad \{\xi<\kappa:f(\xi)=g(\xi)\}\in U^{L[U]}. \end{align*} The membership relation is defined by \begin{align*} [f]\in_{U^{L[U]}}[g] \quad \text{if and only if} \quad \{\xi<\kappa:f(\xi)\in g(\xi)\}\in U^{L[U]}. \end{align*} The two displayed subsets of $\kappa$ are formed inside $L[U]$ by Separation from $f$, $g$, and $\kappa$, and membership in the measure is evaluated using the internal predicate $U^{L[U]}$. So $L[U]$ is $L$ rebuilt with precisely the missing witness added in a controlled way. The prototype is that bare $L$ may contain the ordinal $\kappa$ without containing the measure, while $L[U]$ is designed so that the measure data needed for the ultrapower is part of the model itself. [/example] The chapter's conclusion is that obstacles are informative. The failure of naive measurability in $L$, the covering dichotomy around $0^\sharp$, and the emergence of indiscernibles all point toward the same principle: large cardinals require canonical inner models richer than $L$ but disciplined by the same fine-structural ideals. Chapter 8 begins that replacement by introducing premice, extenders, iterability, and comparison. $L$ exposes the need for inner models that are canonical without being too small. Chapter 8 begins the construction of such models by replacing bare constructibility with premice, extenders, iterability, and comparison. # 8. Inner Model Theory: First Contact Inner model theory asks how much of the universe can be rebuilt from canonical, fine-structural approximations to $V$. Earlier chapters used elementary embeddings and ultrafilters as external evidence for large cardinals; here the direction changes. We try to construct inner models which contain controlled large-cardinal structure, and we ask when those models are correct enough to serve as lower-bound witnesses for consistency strength. The guiding tension is that large cardinals require objects stronger than the definability machinery of $L$, while canonicity demands that these objects be organized in a rigid way. The first inner models beyond $L$ are built from measures and extenders, arranged into structures called mice. This chapter gives a first contact with the vocabulary and the main correctness principle: iterability. ## Premice, Extenders, and the Failure of Measures Alone What should replace $L$ once measurable cardinals enter the picture? In $L$, every stage is obtained by definability over the previous stages, so the construction is canonical but cannot contain a measurable cardinal of the ambient universe. To model measurability internally, we must add coherent ultrafilter-like data to the hierarchy. A single normal measure already suggests the right shape. If $U$ is a normal measure on a measurable cardinal $\kappa$, then an inner model trying to remember $U$ should not merely contain $\kappa$; it should also carry $U$ as part of its structure, since the ultrapower map derived from $U$ is the source of the large-cardinal behaviour. [definition: Premouse] A premouse is a fine-structural extender structure \begin{align*} M=(|M|,\in,E^M,F^M) \end{align*} whose universe $|M|$ is an initial segment of a Jensen-style hierarchy built relative to an amenable extender sequence $E^M$. The predicate $E^M$ indexes extenders over proper initial segments of $M$ in the order prescribed by the hierarchy. The optional active extender $F^M$ is either absent, in which case $M$ is passive, or is an extender over $M$ placed at the top of the structure, in which case $M$ is active. The structure is required to satisfy the usual premouse fine-structural conditions for this course: amenability of the extender predicates, coherence of initial segments, and the rudimentary closure and acceptability properties needed for ultrapowers by extenders on the sequence. [/definition] The word "pre" signals that the structure has the correct formal shape but may not yet be trustworthy. This definition is schematic because full fine structure includes projecta, soundness, standard parameters, and the precise indexing convention for active and passive initial segments. For this chapter, the important correction to the naive picture is that a premouse is not merely a transitive set with a predicate for extenders: the extender sequence must be amenable to the hierarchy and readable by its initial segments. A premouse can carry extenders in the right positions and still behave badly when those extenders are used to form iterated ultrapowers. The distinction between having the syntax of a mouse and being a genuine mouse is made by iterability. Extenders generalize normal measures by coding longer fragments of elementary embeddings. A normal measure on $\kappa$ records the behaviour of an embedding on subsets of $\kappa$; an extender can record how an embedding acts on many finite tuples of ordinals below a larger target. [definition: Extender] Let $M$ be a transitive model of a sufficient fragment of set theory, let $\kappa \in M$ be a cardinal, and let $\lambda$ be an ordinal. A $(\kappa,\lambda)$-extender over $M$ is a system \begin{align*} F=(F_a)_{a\in[\lambda]^{<\omega}} \end{align*} such that, for each finite $a\subset\lambda$ with $n=|a|$, after writing $a=\{a_0<\cdots<a_{n-1}\}$, $F_a$ is a $\kappa$-complete ultrafilter on the Boolean algebra $\mathcal{P}(\kappa^n)\cap M$. The ultrafilters satisfy the projection coherence condition: whenever $a\subset b$ are finite subsets of $\lambda$, membership in $F_a$ is obtained from membership in $F_b$ by pulling back along the coordinate projection from the increasing enumeration of $b$ to the increasing enumeration of $a$. [/definition] This is the simplified course-level version of an extender. More explicitly, if $a\subset b$, $a=\{a_0<\cdots<a_{m-1}\}$, and $b=\{b_0<\cdots<b_{n-1}\}$, then the projection is the map \begin{align*} \pi_{b,a}:\kappa^n\longrightarrow \kappa^m,\qquad \pi_{b,a}(s)(r)=s(q_r), \end{align*} where $b_{q_r}=a_r$. Projection coherence says that, for $X\in \mathcal{P}(\kappa^m)\cap M$, \begin{align*} X\in F_a \iff \pi_{b,a}^{-1}(X)\in F_b. \end{align*} The use of increasing enumerations is only a presentation device: the finite set $a$ is the index, while $\kappa^n$ is the space of functions from the ordered positions of $a$ into $\kappa$. Full extender definitions impose additional coherence, normality, and indexing requirements so that the system is exactly the fragment of an elementary embedding that a fine-structural hierarchy can use. The ordering convention on finite index sets lets the coordinate projections be read unambiguously without replacing finite index sets by arbitrary tuples. The formal system of ultrafilters becomes useful through its associated ultrapower. For this introductory chapter, the following construction records the notation used later without building the full fine-structural apparatus of extender ultrapowers. [definition: Extender Ultrapower] The extender ultrapower $\operatorname{Ult}(M,F)$ has elements represented by pairs $[a,f]_F$, where $a\in[\lambda]^{<\omega}$ and $f\in M$ is a function with domain $\kappa^{|a|}$. The [equivalence relation](/page/Equivalence%20Relation) and membership relation are computed using the ultrafilters $F_a$ and the projection maps between indices. The associated ultrapower map is \begin{align*} j_F:M\longrightarrow \operatorname{Ult}(M,F), \qquad j_F(x)=[\varnothing,c_x]_F, \end{align*} where $c_x$ is the constant function with value $x$. If $\operatorname{Ult}(M,F)$ is well-founded, its transitive collapse is denoted $M_F$, and the collapsed ultrapower embedding is written \begin{align*} j_F:M\longrightarrow M_F. \end{align*} [/definition] When the ultrapower is well-founded, the collapsed embedding $j_F$ has critical point $\kappa$: this means that $\kappa$ is the least ordinal moved by $j_F$. The point of the indexing is to remember more than this single critical point. The finite index $a$ records a finite fragment of the target of the intended embedding, and the projection coherence conditions force these fragments to agree when indices are enlarged or restricted. This extra information is needed when the intended elementary embedding has strength beyond measurability, for instance when it witnesses strongness or Woodinness. Measures are enough for the first canonical model with one measurable cardinal, but they do not give a flexible coding device for stronger embeddings. [example: A Mouse With One Normal Measure] Let $U$ be a normal measure on $\kappa$, and let $M$ be a premouse whose only active extender entry is the measure $U$. The associated ultrapower map is \begin{align*} j_U:M\to \operatorname{Ult}(M,U),\qquad j_U(x)=[c_x]_U, \end{align*} where $c_x(\xi)=x$ for every $\xi<\kappa$. For each ordinal $\alpha<\kappa$, the class $[c_\alpha]_U$ is the same ordinal $\alpha$ in the ultrapower, so $j_U(\alpha)=\alpha$. To see that $\kappa$ is moved, compare $j_U(\kappa)=[c_\kappa]_U$ with the class $[\operatorname{id}]_U$, where $\operatorname{id}(\xi)=\xi$. The ultrapower order gives \begin{align*} [c_\alpha]_U<[\operatorname{id}]_U \iff \{\xi<\kappa:\alpha<\xi\}\in U. \end{align*} The set $\{\xi<\kappa:\alpha<\xi\}$ is a tail of $\kappa$, hence belongs to $U$ because a normal measure contains all co-bounded subsets of $\kappa$. Thus every $\alpha<\kappa$ lies below $[\operatorname{id}]_U$. Also \begin{align*} [\operatorname{id}]_U<[c_\kappa]_U \iff \{\xi<\kappa:\xi<\kappa\}\in U, \end{align*} and the right-hand set is all of $\kappa$, so it is in $U$. Therefore $\kappa\leq[\operatorname{id}]_U<j_U(\kappa)$, and in particular $j_U(\kappa)>\kappa$. This is the first deviation from $L$: the hierarchy is still definable, but the active measure supplies a nonidentity elementary embedding with critical point $\kappa$. [/example] This example is the prototype for mice with finitely or countably many measures. Once extenders are allowed, however, the next problem is coherence: the sequence must not merely list large-cardinal witnesses, but must list them in a way compatible with the initial segments of the model. This motivates the following definition, which isolates the bookkeeping condition needed before iteration can behave canonically. [definition: Jensen Hierarchy Relative To An Extender Sequence] For an extender sequence $E$, the hierarchy $J^E=(J^E_\alpha)_{\alpha\in\operatorname{Ord}}$ is the Jensen-style hierarchy built with $E$ as an amenable predicate. The notation $M\triangleleft J^E_\alpha$ means that $M$ is a proper initial segment of the extender model $J^E_\alpha$. [/definition] The relative hierarchy notation records where each extender is supposed to live. It does not yet control what happens to that location data after an ultrapower: an extender placed at the right initial segment could still send the earlier sequence to a predicate that no longer matches the hierarchy being constructed. Thus the hierarchy needs a compatibility requirement connecting the placement of extenders before the ultrapower with the placement of their images after the ultrapower. This is the role of coherence. The preceding definitions supplied the objects to be arranged: premice, extenders, their ultrapowers, and the relative Jensen hierarchy. The next definition adds the condition that makes those objects usable as one canonical construction rather than as an unrelated list of large-cardinal witnesses. [definition: Coherent Extender Sequence] A coherent extender sequence is a sequence $E$ such that whenever an extender $F$ appears on $E$ over an initial segment $M \triangleleft J^E_\alpha$, the ultrapower map $j_F:M \to \operatorname{Ult}(M,F)$ sends the earlier part $E\upharpoonright M$ of the extender sequence to the corresponding earlier part of the extender sequence of the transitive collapse of $\operatorname{Ult}(M,F)$, whenever this ultrapower is well-founded. [/definition] Coherence is the internal bookkeeping condition that lets the hierarchy recognize its own extender sequence after ultrapowers. Without it, successive ultrapowers would scramble the construction and comparison arguments would lose their canonical target. ## Iterability as the Central Correctness Condition How do we know that a premouse is mathematically usable rather than a formal object carrying inconsistent extender data? The test is not whether the first ultrapower exists; the test is whether all prescribed iteration processes keep producing well-founded models. Inner model theory replaces the simple question "does this model have a measure?" with the more durable question "can this model survive its iterations?" [definition: Iteration of a Premouse] An iteration of a premouse $M$ is a transfinite sequence $(M_i, j_{i,k})_{i \leq k < \theta}$ with $M_0=M$, where each $j_{i,k}:M_i \to M_k$ is an elementary iteration embedding, successor stages are obtained by ultrapowers of current models by extenders on their extender sequences, and at each limit stage $\lambda<\theta$ the model $M_\lambda$ is the transitive collapse of the direct limit of the directed system $(M_i,j_{i,k})_{i\leq k<\lambda}$ whenever that direct limit is well-founded. [/definition] An iteration is a controlled attempt to use the large-cardinal information built into $M$. If the construction ever produces an ill-founded ultrapower, then the earlier extender sequence was not reliable enough to support the intended canonical model. [definition: Iterable Premouse] A premouse $M$ is iterable if every iteration tree on $M$ allowed by the relevant iteration rules has a well-founded cofinal branch and gives rise to well-founded models along that branch. [/definition] This definition suppresses technical choices about the exact class of iteration trees and strategies. In a first course, the important point is that iterability is a global correctness condition, not a local closure property. It says that the extender sequence remains usable under repeated self-application. [definition: Mouse] A mouse is an iterable premouse. [/definition] Thus a mouse is the inner model theorist's replacement for $L$ in the presence of large cardinals: it is fine-structural, canonical, and equipped with enough iteration theory to support comparison. The term is intentionally reserved for structures that have passed this well-foundedness test. The definition is abstract because real iteration trees may branch and may use different extenders at different stages. Before confronting that generality, it is useful to look at the linear case where the same measure is repeatedly moved forward by its own ultrapower embeddings. This isolates the well-foundedness issue without the extra bookkeeping of branch selection. [example: Repeated Ultrapowers by One Measure] Start with a premouse $M_0$ carrying a normal measure $U_0$ on $\kappa_0$. Define \begin{align*} M_1=\operatorname{Ult}(M_0,U_0) \end{align*} and let \begin{align*} j_{0,1}:M_0\to M_1,\qquad j_{0,1}(x)=[c_x]_{U_0}. \end{align*} The next measure is the image \begin{align*} U_1=j_{0,1}(U_0). \end{align*} Since $U_0$ is a normal measure on $\kappa_0$ in $M_0$, elementarity gives that $U_1$ is a normal measure on \begin{align*} \kappa_1=j_{0,1}(\kappa_0) \end{align*} inside $M_1$. Inductively, suppose $M_n$, $U_n$, and $\kappa_n$ have been defined, with $U_n$ a normal measure on $\kappa_n$. Set \begin{align*} M_{n+1}=\operatorname{Ult}(M_n,U_n) \end{align*} and let \begin{align*} j_{n,n+1}:M_n\to M_{n+1},\qquad j_{n,n+1}(x)=[c_x]_{U_n}. \end{align*} For $m<n$, define the iteration embedding by composition: \begin{align*} j_{m,n}=j_{n-1,n}\circ j_{n-2,n-1}\circ\cdots\circ j_{m,m+1}. \end{align*} Then the next critical point is \begin{align*} \kappa_{n+1}=j_{n,n+1}(\kappa_n)=j_{0,n+1}(\kappa_0), \end{align*} because \begin{align*} j_{0,n+1}(\kappa_0)=(j_{n,n+1}\circ j_{0,n})(\kappa_0)=j_{n,n+1}(j_{0,n}(\kappa_0))=j_{n,n+1}(\kappa_n). \end{align*} Likewise the next measure is \begin{align*} U_{n+1}=j_{n,n+1}(U_n). \end{align*} At each successor step, $j_{n,n+1}$ fixes every ordinal below $\kappa_n$ and moves $\kappa_n$. Indeed, if $\alpha<\kappa_n$, then \begin{align*} j_{n,n+1}(\alpha)=[c_\alpha]_{U_n}=\alpha. \end{align*} On the other hand, with $\operatorname{id}_n(\xi)=\xi$ for $\xi<\kappa_n$, the ultrapower order gives \begin{align*} [c_\alpha]_{U_n}<[\operatorname{id}_n]_{U_n}\iff \{\xi<\kappa_n:\alpha<\xi\}\in U_n. \end{align*} The set on the right is a tail of $\kappa_n$, so it belongs to the normal measure $U_n$. Also \begin{align*} [\operatorname{id}_n]_{U_n}<[c_{\kappa_n}]_{U_n}\iff \{\xi<\kappa_n:\xi<\kappa_n\}\in U_n. \end{align*} The right-hand set is $\kappa_n$, hence is in $U_n$. Therefore \begin{align*} \kappa_n\leq[\operatorname{id}_n]_{U_n}<j_{n,n+1}(\kappa_n)=\kappa_{n+1}. \end{align*} Thus the finite iteration is a single linear branch: at stage $n$, one applies the image measure $U_n$ on the image critical point $\kappa_n$, producing the next model $M_{n+1}$ and the next image measure $U_{n+1}$. [/example] The finite stages already reveal why well-foundedness matters. If each $M_n$ is well-founded, then we may collapse it to a transitive model and keep applying the measure sequence. At limit stages, the direct limit of the system must also be well-founded for the iteration to continue as an inner-model construction. The following theorem records the basic criterion suggested by this example. [definition: Countable Linear Iterability For One Measure] Let $M$ be a premouse with one normal measure $U$ on $\kappa$. The pair $(M,U)$ is countably linearly iterable if, for every countable ordinal $\theta$, every sequence $(M_i,U_i,j_{i,k})_{i\leq k<\theta}$ obtained by setting $M_0=M$, $U_0=U$, taking \begin{align*} M_{i+1}=\operatorname{Ult}(M_i,U_i), \qquad U_{i+1}=j_{i,i+1}(U_i) \end{align*} at successor stages, and taking the direct limit at limit stages, has well-founded models at all stages below $\theta$. [/definition] This restricted definition removes the branch-selection problem from general iterability. It tests only the one linear procedure generated by repeatedly moving the same normal measure forward. [quotetheorem:7435] [citeproof:7435] This criterion is much weaker than full iterability for modern mice, but it captures the first reason the definition is necessary. It says only that the single linear procedure generated by $U$ remains well-founded through countable length. The countability assumption matters because the criterion is testing countable iteration systems: a premouse may pass every finite test while failing at the first countable limit if the direct limit has a descending membership sequence. The linearity assumption matters because no branch choice is being made; a branching iteration tree can have two plausible cofinal branches, and the existence of well-founded ultrapowers along one branch does not supply a strategy for choosing a well-founded branch in every tree. The one-measure assumption matters because there is no interaction between overlapping extenders: once two extenders are present, using one can move the critical point, length, or indexing data of the other, and a later ultrapower can fail even if each extender behaves acceptably when used in isolation. A useful boundary case is a countable extender structure with two formally coherent measures $U_0$ and $U_1$ whose individual ultrapowers are well-founded, but where applying $U_0$ first sends the position of $U_1$ to an image measure whose ultrapower is ill-founded. Such a structure can look correct at the level of each separate large-cardinal witness while failing the repeated-use test. Another boundary case is a tree in which one branch repeatedly applies images of $U_0$ and remains well-founded, while a competing branch first applies $U_1$ and reaches an ill-founded limit. These examples explain why full iterability is a strategy property for all relevant trees, not merely a statement about the first ultrapower or about one preferred linear branch. Large-cardinal data is not validated by being written on the sequence; it is validated by its behaviour under iteration. ## Comparison as the Organizing Principle Once we have two iterable premice, how can they be compared? The analogue in $L$ is simple because $L_\alpha$ and $L_\beta$ are linearly ordered by initial segment. Mice need not begin with the same extender sequence, so comparison proceeds by iterating both sides until they line up. [definition: Coiteration] A coiteration of two premice $M$ and $N$ is a simultaneous construction of iteration trees on $M$ and on $N$, choosing extenders according to the first disagreement between the current models, with the aim of producing final models that are comparable by initial segment. [/definition] The first-disagreement rule is the engine of comparison. At any stage, the two current models are inspected until their extender sequences or underlying hierarchies diverge; the side responsible for the disagreement is iterated to remove it. Iterability is what prevents this process from collapsing into ill-foundedness, and the next theorem states the payoff: successful coiteration turns two mice into comparable final models. [explanation: Comparison Principle] In the settings where the course uses comparison, the premice under discussion come with enough iterability to select well-founded branches through the relevant countable iteration trees, and their extender sequences satisfy the fine-structural coherence, soundness, and initial-segment conditions required by the comparison construction. Under those hypotheses, the comparison process builds iteration trees on $M$ and $N$ with last models $M^*$ and $N^*$ such that either $M^*$ is an initial segment of $N^*$ or $N^*$ is an initial segment of $M^*$. [/explanation] The comparison theorem is one of the central technical arguments of inner model theory. Its role here is structural: it explains why iterable premice can be treated as canonical approximations rather than as unrelated models with extender predicates. The hypotheses are doing real work. Without iterability, the construction may reach a limit stage with no well-founded branch; without fine structure, the first-disagreement rule may fail to identify a unique side to move or may destroy the initial-segment relation needed at the end. [remark: Comparison Replaces Absolute Minimality] The constructible universe $L$ is canonical because every level is defined by the same operation and any two levels are immediately comparable. For mice, the extender sequence introduces apparent choices. The Comparison Lemma restores canonicity by showing that iterable choices can be jointly iterated to a common ordering by initial segment. [/remark] Comparison also gives the conceptual reason for iteration strategies. To compare mice in a reproducible way, one needs rules for selecting branches through iteration trees at limit stages. A good strategy is part of the evidence that the mouse is a canonical object rather than merely an iterable structure in isolation. [example: Comparing Two One-Measure Mice] Let $M_0=M$ carry its normal measure $U_0$ on $\kappa_0=\kappa$, and let $N_0=N$ carry its normal measure $W_0$ on $\lambda_0=\lambda$, with $\kappa_0<\lambda_0$. The first disagreement occurs at the smaller measurable cardinal: the $M$-side has measure data at $\kappa_0$, while the $N$-side has not yet reached its measure at $\lambda_0$. Therefore the comparison first forms \begin{align*} M_1=\operatorname{Ult}(M_0,U_0) \end{align*} with embedding \begin{align*} j^M_{0,1}(x)=[c_x]_{U_0}. \end{align*} For every ordinal $\alpha<\kappa_0$, the constant representative gives \begin{align*} j^M_{0,1}(\alpha)=[c_\alpha]_{U_0}=\alpha. \end{align*} To see that the critical point moves, let $\operatorname{id}_0(\xi)=\xi$ for $\xi<\kappa_0$. For each $\alpha<\kappa_0$, \begin{align*} [c_\alpha]_{U_0}<[\operatorname{id}_0]_{U_0}\iff \{\xi<\kappa_0:\alpha<\xi\}\in U_0. \end{align*} The set on the right is a tail of $\kappa_0$, so it belongs to the normal measure $U_0$. Also, \begin{align*} [\operatorname{id}_0]_{U_0}<[c_{\kappa_0}]_{U_0}\iff \{\xi<\kappa_0:\xi<\kappa_0\}\in U_0. \end{align*} The right-hand set is all of $\kappa_0$, hence belongs to $U_0$. Thus \begin{align*} \kappa_0\leq[\operatorname{id}_0]_{U_0}<j^M_{0,1}(\kappa_0). \end{align*} Writing $\kappa_1=j^M_{0,1}(\kappa_0)$ and $U_1=j^M_{0,1}(U_0)$, the $M$-side has been advanced from $\kappa_0$ to $\kappa_1$. If $\kappa_1<\lambda_0$, the smaller disagreement is still on the $M$-side, so the comparison repeats the same step using $U_1$: \begin{align*} M_2=\operatorname{Ult}(M_1,U_1),\qquad \kappa_2=j^M_{1,2}(\kappa_1). \end{align*} At each such stage, $j^M_{n,n+1}$ fixes every ordinal below $\kappa_n$ and moves $\kappa_n$ above itself. Once some current image $\kappa_r$ is no longer below $\lambda_0$, the smaller unresolved measure may be on the $N$-side; then the comparison forms $\operatorname{Ult}(N_0,W_0)$ and replaces $\lambda_0$ by its image. In this way the coiteration alternates only when the ordering of the current image critical points demands it. The point of comparison is not that the original measures agree, but that iterability lets both extender sequences be moved forward until the final models can be tested for an initial-segment relation. [/example] The Comparison Lemma is also the source of uniqueness theorems. When a core model exists at a given large-cardinal level, comparison is what shows that two attempted constructions produce the same canonical model, up to the appropriate notion of isomorphism or initial segment agreement. ## Core Models Below Large-Cardinal Thresholds What happens if the universe has no large cardinal of a specified strength? Inner model theory answers by building the largest canonical mouse compatible with that failure. This object is the core model, usually denoted $K$, and it plays the role of $L$ relative to a large-cardinal threshold. [definition: Core Model Below a Threshold] A core model below a large-cardinal threshold is a canonical iterable extender model $K$ constructed under the hypothesis that no inner model has a large cardinal of that threshold strength. [/definition] The phrase "below a threshold" is essential. There is not a single universal core model covering all possible large cardinals; the construction depends on which large-cardinal strength has been ruled out. Below a measurable cardinal, below a Woodin cardinal, and below stronger hypotheses, the fine structure changes substantially. The course uses the Dodd-Jensen covering theorem below a measurable cardinal as a quoted fine-structural input. In the absence of an inner model with a measurable cardinal, the corresponding core model $K$ covers uncountable sets of ordinals: every uncountable $X\subseteq\operatorname{Ord}$ is contained in some $Y\in K$ with $|Y|=|X|$. This generalizes Jensen's covering phenomenon for $L$: when the relevant large cardinal is absent, the canonical inner model is close to $V$ on uncountable sets of ordinals. [remark: Meaning of Covering] Covering says that $K$ is not merely an inner model with the right fine structure; it approximates the ambient universe well at the level of cardinal arithmetic and sets of ordinals. If covering fails at the corresponding level, that failure is evidence for an inner model with stronger large-cardinal structure. [/remark] The hypothesis excluding an inner measurable is essential. If there is already an inner model with a measurable cardinal, then the Dodd-Jensen core model below a measurable is the wrong object to expect to cover $V$ in this strong way: the measurable supplies extender structure missing from the lower core model, and sets of ordinals can reflect that missing structure. Concretely, forcing or inner-model constructions over a measurable mouse can add uncountable sets of ordinals whose covering behaviour is controlled by the measure sequence rather than by the lower Dodd-Jensen core model, so the lower $K$ no longer has the correct extender information to approximate them. The restriction to uncountable sets is also essential. Countable sets of ordinals need not be contained in a [countable set](/page/Countable%20Set) from $K$; for $L$, adding a Cohen real $r\subseteq \omega$ gives a countable set of ordinals not contained in any countable set of ordinals from $L$, since any such countable $L$-set can be coded in $L$ and would bound the possible values of $r$ too tightly. This is already the distinction between covering for uncountable sets and the much stronger assertion that every real belongs to the inner model. Finally, the theorem is about sets of ordinals rather than arbitrary sets because core model covering is calibrated through ordinal coding and cardinal arithmetic. Thus covering is not a [universal approximation](/theorems/1994) theorem for every universe; it is a theorem about the universe after a particular large-cardinal strength has been ruled out. The contrapositive viewpoint is often the most useful one. Failure of covering can be transformed into consistency strength: if no suitable $K$ can cover the relevant sets, then the universe has enough complexity to imply an inner model with a large cardinal beyond the assumed threshold. This motivates the following organizing principle, which summarizes how core model arguments are used as lower-bound machines. [explanation: Core Model Dichotomy] At a fixed large-cardinal threshold, core model arguments often take the following form: either the corresponding core model $K$ exists and satisfies the expected comparison and covering properties, or the obstruction to constructing such a $K$ yields an inner model with a large cardinal at or above that threshold. This is not a single theorem with one uniform proof across all thresholds. The exact statement depends on which large cardinal is being excluded, which version of the core model is available, and which comparison theorem has been proved for that level of fine structure. Its value in the course is methodological: it explains how failures of canonical approximation become lower bounds in consistency strength. [/explanation] The limitation is important for applications. A failed covering argument does not automatically name the exact large cardinal produced; the surrounding fine-structural theorem determines the threshold. This is why inner model theory interacts so strongly with forcing and descriptive set theory: forcing arguments often create failures of covering or reflection patterns, while determinacy and regularity phenomena provide targets whose consistency strength is calibrated by the appropriate core model dichotomy. For example, a forcing extension may preserve cardinals while adding a set of ordinals whose approximation properties cannot be matched by the old core model; the inner-model question is then whether this failure reflects a missing extender. In descriptive set theory, regularity properties and projective determinacy are compared against mice with Woodin cardinals, so the same comparison-and-covering technology becomes a calibration tool rather than only a construction of $K$. The chapter's earlier focus on iterability is what makes these applications possible: without iterable mice, the lower-bound argument has no canonical model to compare with the forcing or determinacy phenomenon. [example: Covering Below a Measurable Cardinal] Assume there is no inner model with a measurable cardinal, and let $A\subseteq\omega_1$ be uncountable. Since $A\subseteq\omega_1$, every element of $A$ is an ordinal, so $A$ is a set of ordinals. Since $A$ is uncountable and $A\subseteq\omega_1$, its cardinality in $V$ is forced to be $\omega_1$: it cannot be finite or countable, and it cannot exceed $\omega_1$. Thus \begin{align*} |A|^V=\omega_1. \end{align*} By the the covering lemma above, applied to the uncountable set of ordinals $A$, there is a set $B\in K$ of ordinals such that \begin{align*} A\subseteq B. \end{align*} The same theorem also gives \begin{align*} |B|^V=|A|^V. \end{align*} Combining this with $|A|^V=\omega_1$ yields \begin{align*} |B|^V=\omega_1. \end{align*} So $K$ contains an $\omega_1$-sized set of ordinals covering $A$, even though the theorem does not imply that $A\in K$. This is the sense in which $K$ stays close to $V$: it supplies correctly sized approximations to uncountable sets of ordinals without having to contain all of those sets. [/example] This first contact with inner model theory should leave three ideas in place. Premice are fine-structural hierarchies carrying extender data; mice are the iterable premice; and comparison is the theorem that makes mice canonical enough to organize consistency strength. The next chapter returns to the determinacy consequences introduced with Woodin cardinals: the large-cardinal hierarchy around sets of reals depends on refining comparison and iterability rather than replacing them. Inner model theory supplies the canonical machinery needed to organize large-cardinal strength. The next chapter returns to sets of reals, where determinacy turns that strength into regularity phenomena for infinite games and definable subsets of the continuum. # 9. Determinacy and Regularity Properties Determinacy shifts the study of sets of reals from well-ordering and choice toward infinite games. Earlier chapters used large cardinals to measure consistency strength through embeddings and inner models; this chapter explains one of the central reasons Woodin cardinals matter: they are tied to determinacy principles for definable sets of reals. The guiding question is how much regularity follows if every sufficiently definable game has a winning strategy. ## Infinite Games on Omega and the Axiom of Determinacy The first problem is to turn a set of reals into a mathematical contest whose outcome records membership in that set. Infinite games provide exactly this translation: two players build a real number one natural number at a time, and a payoff set decides who wins. We begin by fixing the game form, because later determinacy axioms quantify over this exact object. [definition: Gale Stewart Game] Let $A \subseteq \mathbb N^{\mathbb N}$. The Gale-Stewart game $G(A)$ is the length-$\omega$ game in which players I and II alternately choose natural numbers \begin{align*} n_0, n_1, n_2, \dots . \end{align*} Player I chooses $n_{2k}$ and player II chooses $n_{2k+1}$. The subscripts $0,1,2,\dots$ index turns; the chosen values themselves lie in $\mathbb N$, with the standing notes convention that $\mathbb N$ starts at $1$. The resulting sequence $x \in \mathbb N^{\mathbb N}$ is defined by $x(k)=n_k$. Player I wins iff $x \in A$; otherwise player II wins. [/definition] A payoff set is therefore a set of possible completed plays. To ask whether a player can force a win, we need to formalise rules that depend only on the finite history already played. [definition: Strategy] A strategy for player I in $G(A)$ is a function \begin{align*} \sigma : \mathbb N^{<\mathbb N}_{\mathrm{even}} \to \mathbb N, \end{align*} where $\mathbb N^{<\mathbb N}_{\mathrm{even}}$ is the set of finite sequences $s\in\mathbb N^{<\mathbb N}$ whose length is even. A strategy for player II is a function \begin{align*} \tau : \mathbb N^{<\mathbb N}_{\mathrm{odd}} \to \mathbb N, \end{align*} where $\mathbb N^{<\mathbb N}_{\mathrm{odd}}$ is the set of finite sequences $s\in\mathbb N^{<\mathbb N}$ whose length is odd. [/definition] The strategy convention records the finite play seen by the player whose turn it is. The next notion isolates the situation in which the game has no strategic ambiguity: one of the two players can force the outcome no matter how the opponent responds. [definition: Determined Game] The game $G(A)$ is determined if either player I has a winning strategy in $G(A)$ or player II has a winning strategy in $G(A)$. [/definition] Determined games are individual objects, and ZFC proves determinacy only for many structured classes of payoff sets, not for arbitrary $A\subseteq\mathbb N^{\mathbb N}$. The obstruction is that an arbitrary payoff set may be too wild for the usual definable-game arguments, so one can ask for determinacy as a global replacement principle. The axiom below asserts that the strategic dichotomy holds for every payoff set of reals. [definition: Axiom of Determinacy] The axiom of determinacy, denoted $\mathrm{AD}$, is the assertion that for every $A \subseteq \mathbb N^{\mathbb N}$, the game $G(A)$ is determined. [/definition] The symbol should be read as a schema over all sets of reals. It is not a theorem of $\mathrm{ZFC}$; it conflicts with the full axiom of choice, but it produces a highly regular theory of the real line. [example: Basic Gale Stewart Payoff] Let $A=\{x\in\mathbb N^{\mathbb N}:x(0)=1\}$. Define a strategy $\sigma$ for player I by setting $\sigma(\varnothing)=1$; on later even-length finite histories, $\sigma$ may choose any natural number, for example $\sigma(s)=1$ for every even-length $s$. For any play following $\sigma$, the first move is $n_0=\sigma(\varnothing)=1$. The completed sequence $x\in\mathbb N^{\mathbb N}$ is defined by $x(k)=n_k$, so \begin{align*} x(0)=n_0=1. \end{align*} Hence $x\in A$, independent of the values of $n_1,n_2,n_3,\dots$. Thus $\sigma$ is a winning strategy for player I, and this payoff is decided entirely by the first coordinate of the completed play. [/example] Finite-dependence examples are useful for calibration, but the point of determinacy is that payoffs may depend on the entire infinite sequence. This motivates the closed-game definition: it is the first topological condition under which infinite dependence can still be controlled by finite evidence. [definition: Closed Game] A Gale-Stewart game $G(A)$ is closed for player I if $A \subseteq \mathbb N^{\mathbb N}$ is closed in the [product topology](/page/Product%20Topology) on Baire space. [/definition] Closedness means that if player I loses, this failure is witnessed after finite time: the completed play lies outside $A$, and the complement of $A$ is open. This motivates a determinacy theorem proved by ranking finite positions in the associated tree. [quotetheorem:7436] [citeproof:7436] This result already indicates the shape of later determinacy proofs. A topological condition on the payoff set is converted into a structural property of a tree, and the absence of an infinite descending ordinal sequence supplies the winning-strategy dichotomy. The closedness hypothesis is doing real work for the proof as stated: if the payoff is not closed, a player may only learn at the end of the play whether the target condition was met, so the finite-rank argument has no well-founded tree on which to stop. For example, the set of sequences with infinitely many nonzero entries is not closed; every finite position is compatible both with eventual membership and with eventual failure. The stated theorem directly covers closed payoffs for player I, while open payoffs follow at once by applying the theorem to the closed complement and swapping the players. Countable unions of closed payoffs and general Borel payoffs require the stronger Borel determinacy theorem below. [example: A Closed Payoff] Let $T\subseteq \mathbb N^{<\mathbb N}$ be a tree, meaning that whenever $t\in T$ and $s$ is an initial segment of $t$, then $s\in T$. Its body is \begin{align*} [T]=\{x\in\mathbb N^{\mathbb N}: x{\upharpoonright}k\in T \text{ for every } k\in\mathbb N\}. \end{align*} We show that $A=[T]$ is closed. If $x\notin [T]$, then by the displayed definition there is some $k\in\mathbb N$ such that $x{\upharpoonright}k\notin T$. Let $s=x{\upharpoonright}k$. Every $y\in\mathbb N^{\mathbb N}$ extending $s$ satisfies \begin{align*} y{\upharpoonright}k=s=x{\upharpoonright}k\notin T. \end{align*} Therefore no such $y$ lies in $[T]$, so the basic open neighbourhood $\{y\in\mathbb N^{\mathbb N}: y{\upharpoonright}k=s\}$ is contained in $\mathbb N^{\mathbb N}\setminus [T]$. Thus the complement of $[T]$ is open, and hence $[T]$ is closed. By the closed determinacy theorem above, the game $G([T])$ is determined. Player I wins exactly when the completed play $x$ satisfies $x{\upharpoonright}k\in T$ for every $k$, so player I can force the play to remain on $T$ forever; otherwise player II has a winning strategy, and the completed play must have some finite initial segment outside $T$. In this example, closedness is precisely the fact that losing membership in $[T]$ is witnessed by a finite position. [/example] ## Borel Determinacy and Projective Determinacy Closed sets are only the first level of the Borel hierarchy, so the next question is how far the rank argument can be pushed. Borel sets are made from open sets by countable unions and complements, and the major theorem is that these operations still preserve determinacy. [definition: Borel Determinacy] Borel determinacy is the assertion that $G(A)$ is determined for every Borel set $A \subseteq \mathbb N^{\mathbb N}$. [/definition] The definition is short, but the theorem behind it is deep because Borel sets are built by countable iteration of complementation and countable union. Determinacy must be preserved through this entire hierarchy, and Martin's theorem supplies that preservation inside $\mathrm{ZFC}$. [quotetheorem:7437] Martin's theorem is a fundamental background result for the rest of the chapter. Its proof uses induction through Borel ranks, auxiliary games, and strategy-transfer arguments, and these techniques are usually developed in a separate descriptive set theory course. The point is not just that open and closed games have already been handled. A Borel set may be obtained by iterating complements and countable unions through any countable ordinal rank, and a strategy for a countable union cannot usually be assembled by selecting a winning strategy for one visible closed component. The proof must therefore track the Borel code itself, transfer strategies between the original game and auxiliary rank-lowering games, and show that determinacy survives each countable-stage operation. The scope of the theorem is also important: it proves determinacy exactly for Borel payoff sets inside $\mathrm{ZFC}$, not for arbitrary subsets of $\mathbb N^{\mathbb N}$. The restriction is necessary. Using a well-ordering of the reals, choice can build payoff sets whose associated games are not determined: a standard diagonal construction enumerates all strategies for both players and chooses one completed play defeating each strategy at its assigned stage. Such a payoff cannot be Borel, since Martin's theorem would determine it. The obstruction is therefore not a missing technical refinement of the Borel-rank induction; arbitrary sets of reals can encode a global choice-based diagonalisation against all strategies. Stronger hypotheses such as projective determinacy or full determinacy are needed only after the payoff class is enlarged in a controlled way. [example: Countable Boolean Combinations of Closed Payoffs] Let $(F_m)_{m\in\mathbb N}$ be closed subsets of $\mathbb N^{\mathbb N}$, and define \begin{align*} A = \bigcup_{m=1}^{\infty} F_m. \end{align*} For each completed play $x\in\mathbb N^{\mathbb N}$, the displayed definition gives \begin{align*} x\in A \quad\text{if and only if}\quad \text{there exists }m\in\mathbb N\text{ such that }x\in F_m. \end{align*} Thus $A$ is a countable union of closed sets, so $A$ is an $F_\sigma$ subset of $\mathbb N^{\mathbb N}$. Every [closed set](/page/Closed%20Set) is Borel, and the Borel sets are closed under countable unions; hence $A$ is Borel. By Martin's Borel determinacy theorem above, every Borel payoff set in $\mathbb N^{\mathbb N}$ is determined, so $G(A)$ is determined. The conclusion is stronger than closed determinacy alone: a play is winning for player I when it eventually belongs to at least one of the closed targets $F_m$, but the index $m$ need not be fixed in advance by a single closed tree controlling all winning plays. [/example] Borel determinacy is provable in $\mathrm{ZFC}$, but the next natural definability class already points toward large cardinals. Projective sets arise when Borel relations are allowed to quantify over auxiliary reals, so we need a definition that records closure under this kind of projection. [definition: Projective Set] A set $A \subseteq \mathbb N^{\mathbb N}$ is projective if it belongs to the smallest class of subsets of Baire space that contains all Borel sets and is closed under complementation and continuous images. [/definition] Projection is the source of the jump in strength. It changes questions about a single real into questions about the existence of another real witnessing a relation. This motivates the corresponding determinacy principle for the whole projective hierarchy. [definition: Projective Determinacy] Projective determinacy, denoted $\mathrm{PD}$, is the assertion that $G(A)$ is determined for every projective set $A \subseteq \mathbb N^{\mathbb N}$. [/definition] Projective determinacy is not proved in $\mathrm{ZFC}$. Since projective games encode quantification over reals, their determinacy becomes a test case for the large-cardinal strength developed earlier in the course, motivating the following large-cardinal implication. [quotetheorem:7438] This theorem is quoted as part of the large-cardinal landscape. Its proof uses inner model theory, generic absoluteness, and iteration strategies, and the statement is the part needed here: Woodin cardinals provide the strength behind projective regularity. The large-cardinal assumption is not cosmetic. In Gödel's constructible universe $L$, there is a projective well-ordering of the reals; from such a definable well-ordering one obtains projective regularity failures, for example projective sets without the perfect set property, and hence projective determinacy fails. Thus $\mathrm{ZFC}$ alone cannot prove $\mathrm{PD}$, because $\mathrm{ZFC}$ has models such as $L$ in which a definable choice principle produces projective pathologies. Conversely, the consistency strength of broad projective determinacy is measured by Woodin-cardinal hypotheses, so the connection runs both ways: large cardinals imply determinacy, and determinacy principles reflect substantial large-cardinal strength. The conclusion is projective determinacy, not the full axiom of determinacy. Thus it applies to sets obtained from Borel sets by finitely many real quantifiers and complements, but it does not decide games whose payoff is an arbitrary subset of Baire space unless additional determinacy strength is assumed. [example: A Projective Payoff Set] Let $B \subseteq \mathbb N^{\mathbb N}\times\mathbb N^{\mathbb N}$ be Borel, and define \begin{align*} A = \{x \in \mathbb N^{\mathbb N} : \exists y \in \mathbb N^{\mathbb N}\ ((x,y)\in B)\}. \end{align*} This definition says exactly that $A$ is the projection of $B$ onto the first coordinate. More explicitly, for each $x\in\mathbb N^{\mathbb N}$, \begin{align*} x\in A \quad\text{if and only if}\quad \text{there is }y\in\mathbb N^{\mathbb N}\text{ such that }(x,y)\in B. \end{align*} Since analytic sets are precisely projections of Borel subsets of $\mathbb N^{\mathbb N}\times\mathbb N^{\mathbb N}$, the set $A$ is analytic. Analytic sets are projective because the projective sets contain all Borel sets and are closed under continuous images, and the first-coordinate projection \begin{align*} \pi_1:\mathbb N^{\mathbb N}\times\mathbb N^{\mathbb N}\to\mathbb N^{\mathbb N},\qquad \pi_1(x,y)=x \end{align*} is continuous. Thus \begin{align*} A=\pi_1[B], \end{align*} so $A$ is projective. The game $G(A)$ is therefore the game in which the players build a real $x\in\mathbb N^{\mathbb N}$, and player I wins exactly when some real $y$ witnesses $(x,y)\in B$. Projective determinacy asserts that every projective payoff set is determined, so under $\mathrm{PD}$ this projected Borel payoff still has a winning strategy for one of the two players. [/example] The example shows why projective determinacy is a natural bridge between descriptive set theory and large cardinals. A strategy for such a game must control not only the visible play, but also the possible real witnesses encoded by projection. ## Regularity Consequences for Sets of Reals The next problem is to understand why determinacy is valuable. Its power is that winning strategies impose structure on arbitrary sets of reals, leading to the regularity properties that choice alone cannot guarantee. We first examine the size-theoretic regularity property that rules out arbitrary intermediate subsets of the continuum. [definition: Perfect Set Property] A set $A \subseteq \mathbb R$ has the perfect set property if either $A$ is countable or there is a nonempty perfect set $P \subseteq A$. [/definition] A perfect subset of the real line has cardinality $2^{\aleph_0}$. Thus the perfect set property rules out intermediate-size counterexamples inside a single set of reals, and this motivates the determinacy theorem that gives the dichotomy for every set of reals. [quotetheorem:7439] [citeproof:7439] This theorem gives a sharp contrast with $\mathrm{AC}$. A well-ordering of the reals supports the usual Bernstein construction: an uncountable set $B \subseteq \mathbb R$ such that both $B$ and $\mathbb R\setminus B$ meet every nonempty perfect set. Such a set is uncountable but contains no nonempty perfect subset, so it fails the perfect set property. Thus the theorem is not a statement about arbitrary models with choice; it uses the full force of determinacy for all sets of reals. What it gives is a dichotomy internal to $\mathrm{ZF}+\mathrm{AD}$: every set of reals is either small in the countable sense or already contains a full perfect continuum of points. The next regularity property concerns category rather than cardinality: instead of asking whether a set contains a perfect subset, we ask whether it differs from an open set only by a meagre set. [definition: Baire Property] A set $A \subseteq \mathbb R$ has the Baire property if there is an open set $U \subseteq \mathbb R$ such that $A \triangle U$ is meagre. [/definition] The Baire property says that, modulo a small topological error, the set behaves like an open set. For arbitrary sets of reals, the obstruction is that there may be no definable approximation process producing such an open set, and under choice one can build sets that evade all category approximations. Determinacy changes the situation by turning category questions into games where a winning strategy supplies the required local control. [quotetheorem:7440] [citeproof:7440] The Baire-property argument illustrates a recurring theme: a regularity property follows when every local game has a winner and the relevant topology has a countable basis. It also explains why determinacy interacts naturally with ordinary analysis. Many functional-analytic arguments use the [Baire category theorem](/theorems/630) to turn countably many dense open requirements into a single point satisfying all of them; the proof above is the set-theoretic analogue, with a winning strategy replacing the analytic construction of dense open sets. Under determinacy, arbitrary sets of reals become compatible with this category calculus, so category arguments no longer break down at the moment a non-definable target set appears. The hypothesis $\mathrm{AD}$ is essential here because full choice permits sets with no Baire approximation. A standard Bernstein set gives a clean obstruction. Such a set $B\subseteq\mathbb R$ meets every nonempty perfect set, and so does its complement. If $B$ were meagre on some nonempty open interval, the complement of that meagre set would contain a nonempty perfect set disjoint from $B$. If $B$ were comeagre on some nonempty open interval, the same argument applied to $\mathbb R\setminus B$ would produce a nonempty perfect set disjoint from the complement. Thus a Bernstein set cannot agree with an open set modulo a meagre set. The theorem therefore does not say that every set of reals has the Baire property in $\mathrm{ZFC}$; in that setting the conclusion holds for Borel and many definable sets, but not for arbitrary sets. The measure-theoretic version asks for approximation by Borel sets up to null error, so it requires a parallel definition using Lebesgue measure. [definition: Lebesgue Measurability] A set $A \subseteq \mathbb R$ is Lebesgue measurable if there is a Borel set $B \subseteq \mathbb R$ such that $A \triangle B$ has Lebesgue measure $0$. [/definition] Measure regularity is subtler than category regularity, but it follows from the same strategic philosophy: a player tries to localise the real while the other tries to force the outcome into or out of the target set. This motivates the theorem completing the standard trio of regularity consequences. [quotetheorem:7441] [citeproof:7441] Together, the perfect set property, Baire property, and Lebesgue measurability explain why determinacy is a regularity axiom. The measurability theorem uses $\mathrm{AD}$ for the specific games attached to arbitrary sets $A\subseteq\mathbb R$; weaker determinacy hypotheses give correspondingly weaker conclusions. For example, $\mathrm{PD}$ yields measurability for projective sets by applying the same [regularity theorem](/theorems/2750) only to projective payoff games, while Borel determinacy alone is enough for Borel sets but says nothing about a non-definable subset of $\mathbb R$. The theorem also depends on rejecting the full choice principles used to build pathological sets. Under $\mathrm{AC}$, a Vitali set is obtained by choosing one representative from each coset of $\mathbb R/\mathbb Q$; [translation invariance](/theorems/4911) and countable additivity then show that it cannot be Lebesgue measurable. Hence $\mathrm{AD}$ is not merely proving a difficult theorem of analysis inside $\mathrm{ZFC}$; it is replacing the choice-based universe of arbitrary representatives with a game-theoretic universe in which every set of reals admits a Borel measurable envelope up to null error. [example: Regularity Under Projective Determinacy] Assume $\mathrm{PD}$, and let $A\subseteq\mathbb R$ be projective. Using the fixed coding of reals by elements of Baire space, let $\widetilde A\subseteq\mathbb N^{\mathbb N}$ be the corresponding payoff set. Projectiveness is preserved by this coding because the coding maps used are Borel, hence projective, and the projective sets are closed under complements and continuous images. Therefore $\widetilde A$ is projective. By $\mathrm{PD}$, the Gale-Stewart game $G(\widetilde A)$ is determined. The proofs of the determinacy regularity theorems above use determinacy only for the payoff games attached to the set under consideration. Applying those arguments to the projective payoff $\widetilde A$ gives that $A$ is Lebesgue measurable, has the Baire property, and has the perfect set property. For the analytic case, if $B\subseteq\mathbb N^{\mathbb N}\times\mathbb N^{\mathbb N}$ is Borel and \begin{align*} A=\{x\in\mathbb N^{\mathbb N}:\exists y\in\mathbb N^{\mathbb N}\ ((x,y)\in B)\}, \end{align*} then $A=\pi_1[B]$ for the continuous projection $\pi_1(x,y)=x$. Since Borel sets are projective and projective sets are closed under continuous images, this $A$ is projective. Thus under $\mathrm{PD}$, projecting a Borel relation may add real quantifiers, but it still cannot produce the classical nonmeasurable, non-Baire, or perfect-set-pathological examples. [/example] ## Determinacy Versus Choice The final problem is to reconcile determinacy with the axiom of choice. The answer is that they cannot coexist in full strength: determinacy gives regularity for all sets of reals, while choice gives well-orderings that generate non-regular sets. The incompatibility is formalised by comparing the regularity theorems above with the usual choice-based constructions. [quotetheorem:7442] [citeproof:7442] The contradiction uses full choice through the global well-ordering and representative-selection arguments that build Vitali and Bernstein-type pathologies. It does not rule out every choice principle. Countable recursive choices are much weaker: they ask for a sequence chosen step by step from a serial relation, not for a simultaneous choice from an arbitrary family or for a well-ordering of $\mathbb R$. This distinction is essential because the regularity proofs above themselves use countable constructions of plays, trees, bases, and approximations. Since analysis still needs these countable recursive constructions, the useful question is which fragments of choice survive alongside determinacy. [definition: Dependent Choice] The axiom of dependent choice, denoted $\mathrm{DC}$, states that for every nonempty set $X$ and every binary relation $R$ on $X$ such that for every $x\in X$ there is $y\in X$ with $xRy$, there exists a sequence $(x_n)_{n\in\mathbb N}$ in $X$ such that $x_n R x_{n+1}$ for all $n\in\mathbb N$. [/definition] Dependent choice is strong enough for most countable constructions in analysis, including the recursive construction of sequences. It is weak enough to be compatible with determinacy in the usual determinacy models. [remark: Choice Fragments in Determinacy Models] The standard background theory for determinacy is often $\mathrm{ZF}+\mathrm{DC}+\mathrm{AD}$ rather than $\mathrm{ZFC}+\mathrm{AD}$. This keeps the countable choice principles needed for analysis while rejecting the global well-ordering of the reals. [/remark] The large-cardinal significance of the chapter is now visible. Woodin cardinals provide consistency strength for determinacy statements about definable sets of reals, and determinacy translates that strength into concrete regularity theorems. In this way, large cardinals do not merely assert the existence of very large infinities; they organise the fine structure of the continuum. Determinacy shows that large cardinals have concrete consequences far below their own size, especially in the structure of the real line. The next chapter develops this connection systematically through projective sets, scales, and absoluteness principles in descriptive set theory. # 10. Large Cardinals and Descriptive Set Theory This chapter connects the large-cardinal hierarchy with the fine structure of definable sets of reals. Earlier chapters introduced Woodin cardinals through elementary embeddings, extenders, and consistency strength; here the same hypotheses appear as regularity principles for projective sets and as absoluteness principles for statements about reals. The guiding question is how much set-theoretic strength is needed before definable subsets of Polish spaces behave like Borel sets rather than like arbitrary subsets of the continuum. ## Projective Sets, Games, and Scales Which definable sets of reals can be studied without naming arbitrary subsets of the continuum? Descriptive set theory begins with Borel sets, then closes under projection and complementation to obtain the projective hierarchy. The hierarchy is robust enough to include many naturally occurring sets, but complex enough that ZFC alone cannot settle all of its regularity questions. [definition: Projective Hierarchy] Let $X$ be a Polish space. The projective pointclasses on $X$ are defined recursively as follows: 1. $\mathbf{\Sigma}^1_1(X)$ is the class of sets $A\subseteq X$ for which there are a Polish space $Y$, a Borel set $B\subseteq Y$, and a continuous map $f:Y\to X$ such that $A=f(B)$. 2. $\mathbf{\Pi}^1_n(X)=\{X\setminus A : A\in \mathbf{\Sigma}^1_n(X)\}$. 3. $\mathbf{\Sigma}^1_{n+1}(X)$ is the class of projections of sets in $\mathbf{\Pi}^1_n(X\times \omega^\omega)$. 4. $\mathbf{\Delta}^1_n(X)=\mathbf{\Sigma}^1_n(X)\cap \mathbf{\Pi}^1_n(X)$. [/definition] Thus analytic sets are the first projective sets, coanalytic sets are their complements, and higher levels alternate projection and complementation. The lightface classes $\Sigma^1_n$, $\Pi^1_n$, and $\Delta^1_n$ refer to definitions without arbitrary real parameters; boldface notation allows real parameters. [example: Analytic Ill-Founded Trees] Let $\operatorname{Tr}$ be the Polish space of trees on $\omega$, viewed as subsets of $\omega^{<\omega}$. Define \begin{align*} IF=\{T\in \operatorname{Tr}: T \text{ has an infinite branch}\}. \end{align*} We show that $IF$ is analytic by writing it as the projection of a closed set in $\operatorname{Tr}\times \omega^\omega$. For $T\in\operatorname{Tr}$ and $x\in\omega^\omega$, let $x|n$ denote the finite sequence $\langle x(0),\ldots,x(n-1)\rangle$. Consider \begin{align*} B=\{(T,x)\in \operatorname{Tr}\times \omega^\omega : \forall n\in\mathbb N,\ x|n\in T\}. \end{align*} For each fixed $n$, the condition $x|n\in T$ depends only on the finite initial segment $x|n$ and on whether the coordinate $x|n$ belongs to the subset $T\subseteq\omega^{<\omega}$. Hence the set \begin{align*} B_n=\{(T,x):x|n\in T\} \end{align*} is clopen in $\operatorname{Tr}\times\omega^\omega$. Since \begin{align*} B=\bigcap_{n\in\mathbb N}B_n, \end{align*} the set $B$ is closed. Now compute the projection of $B$ to the first coordinate: \begin{align*} \operatorname{proj}_{\operatorname{Tr}}(B)=\{T\in\operatorname{Tr}:\exists x\in\omega^\omega\ (T,x)\in B\}. \end{align*} By the definition of $B$, this is \begin{align*} \operatorname{proj}_{\operatorname{Tr}}(B)=\{T\in\operatorname{Tr}:\exists x\in\omega^\omega\ \forall n\in\mathbb N,\ x|n\in T\}. \end{align*} The condition $\forall n,\ x|n\in T$ says exactly that $x$ is an infinite branch through $T$, so \begin{align*} \operatorname{proj}_{\operatorname{Tr}}(B)=IF. \end{align*} Thus $IF$ is analytic, because it is the projection of a closed, hence Borel, subset of $\operatorname{Tr}\times\omega^\omega$. Its complement $WF=\operatorname{Tr}\setminus IF$ is therefore coanalytic. This example shows why projection is the natural operation at the first projective level: the tree $T$ is the object being classified, while the infinite branch $x$ is an existential witness living in a second copy of Baire space. [/example] The tree example separates the well-founded case, controlled by ranks, from the ill-founded case, controlled by branches. To use this rank-versus-branch analysis beyond analytic sets, we need a systematic way to attach ordinal ranks to members of a definable set while preserving limits of approximating reals. [definition: Scale] Let $A\subseteq \omega^\omega$. A scale on $A$ is a sequence $(\varphi_n)_{n\in \mathbb N}$ of norms $\varphi_n:A\to \lambda_n$, where each $\lambda_n$ is an ordinal, such that whenever $(x_k)_{k\in \mathbb N}$ is a sequence in $A$ converging to $x\in \omega^\omega$ and for each $n$ the ordinal sequence $(\varphi_n(x_k))_{k\in \mathbb N}$ is eventually constant with value $\alpha_n$, then $x\in A$ and $\varphi_n(x)\le \alpha_n$ for every $n$. [/definition] A scale turns a set of reals into a ranked object whose limits can be controlled. The complementary method is strategic rather than rank-theoretic: to connect large cardinals with projective regularity, we need to encode a set of reals as the payoff set of an infinite two-player contest. [definition: Infinite Game on Reals] For $A\subseteq \omega^\omega$, the game $G_A$ is the same Gale-Stewart game from Chapter 9, now written in the standard descriptive-set-theoretic notation: two players alternately play natural numbers, producing $x\in \omega^\omega$. Player I wins iff $x\in A$. [/definition] The game formulation turns membership in a set of reals into a strategic question. Regularity properties follow because a winning strategy gives a uniform way to build or avoid points in the set. [definition: Determined Set] A set $A\subseteq \omega^\omega$ is determined if one of the two players has a winning strategy in $G_A$. [/definition] Determinacy is incompatible with the full axiom of choice when asserted for all sets of reals, but it is consistent to ask for determinacy only for definable pointclasses. Projective determinacy is the assertion that every projective subset of $\omega^\omega$ is determined. [example: Closed Games] Let $A\subseteq\omega^\omega$ be closed. Since $\omega^\omega\setminus A$ is open, define the tree of finite positions compatible with $A$ by \begin{align*} T_A=\{s\in\omega^{<\omega}:\exists x\in A\text{ such that }s\subseteq x\}. \end{align*} Then \begin{align*} A=\{x\in\omega^\omega:\forall n\in\mathbb N,\ x|n\in T_A\}. \end{align*} Indeed, if $x\in A$, every initial segment of $x$ lies in $T_A$ by taking $x$ itself as the witness. Conversely, if $x\notin A$, then $x\in\omega^\omega\setminus A$, so by openness there is some $n$ such that every real extending $x|n$ lies outside $A$; hence $x|n\notin T_A$. Call a finite position $s$ good if Player I can force all later finite positions to remain in $T_A$ starting from $s$. This condition is determined recursively by the player to move. If $s\notin T_A$, then $s$ is not good. If $s\in T_A$ and it is Player I's turn, then $s$ is good exactly when there is some $k\in\omega$ such that $s^\frown\langle k\rangle$ is good. If $s\in T_A$ and it is Player II's turn, then $s$ is good exactly when for every $k\in\omega$, the position $s^\frown\langle k\rangle$ is good. If the empty position $\varnothing$ is good, Player I chooses at each of his turns a number $k$ for which the next position remains good. Player II's moves also preserve goodness by the defining universal clause for Player II turns. Therefore every finite initial segment of the resulting play lies in $T_A$, so the produced real lies in $A$ by the displayed characterization of $A$. If $\varnothing$ is not good, Player II instead keeps the play at not-good positions whenever it is his turn. When a not-good position $s\in T_A$ is on Player II's turn, the recursive condition says that some $k$ has $s^\frown\langle k\rangle$ not good, and Player II plays such a $k$. When it is Player I's turn at a not-good position, every legal next position is not good; otherwise $s$ would be good by the existential clause. Thus Player I cannot force the play to remain inside $T_A$ forever. Since $A$ consists exactly of the branches all of whose finite initial segments lie in $T_A$, Player II's strategy produces a real outside $A$. Hence $G_A$ is determined. The proof shows the basic rank-versus-tree pattern: closed payoff sets are controlled by finite positions, and the losing player is forced out of the tree of positions compatible with the payoff. [/example] ## Shoenfield Absoluteness as the Baseline What can ZFC already say about projective truth when we pass to forcing extensions or transitive models? The first benchmark is Shoenfield absoluteness. It says that the second projective level is too low to be changed by forcing, provided the parameters remain fixed. [definition: Shoenfield Statement] A statement about a real parameter $a\in \omega^\omega$ is Shoenfield if it has complexity $\Sigma^1_2(a)$ or $\Pi^1_2(a)$ over the real universe. [/definition] This terminology isolates the part of the projective hierarchy where absoluteness follows from countable-model arguments rather than from large cardinals. The possible obstruction is that moving to a forcing extension can add new reals, new branches through coded trees, and new witnesses to projective statements. Shoenfield absoluteness identifies the low-complexity boundary where those new witnesses cannot change truth for fixed real parameters. [quotetheorem:7443] [citeproof:7443] The hypotheses are doing real work. Transitivity ensures that membership in $\omega$, finite sequences, trees, and branches is interpreted correctly; without it, a model may believe a coded tree is well-founded while the ambient universe sees a descending branch. Having the same ordinals prevents well-founded ranks from being shortened or extended when moving between the models. The fixed real parameter condition matters because forcing may add a new parameter $b$ and then make new low-complexity statements about $b$ true; Shoenfield absoluteness does not say that truth is absolute uniformly for parameters that do not exist in the ground model. The theorem is therefore a precise baseline rather than a blanket invariance principle: beyond $\Sigma^1_2$ and $\Pi^1_2$, forcing can affect projective truth unless stronger structure, such as determinacy or Woodin-cardinal generic absoluteness, is available. [example: Sigma One Two Absoluteness] Let $P\subseteq \omega^\omega\times \omega^\omega\times \omega^\omega$ be arithmetical, and fix a real parameter $a\in V\cap\omega^\omega$. Consider the statement \begin{align*} \exists x\in \omega^\omega\,\forall y\in \omega^\omega\, P(x,y,a). \end{align*} Because $P(x,y,a)$ is arithmetical, it has only quantifiers over natural numbers after the real variables $x,y,a$ are fixed. Therefore the displayed formula has the projective form “there exists a real $x$ such that for every real $y$ an arithmetical condition holds,” so it is a $\Sigma^1_2(a)$ statement. Now let $G$ be $V$-generic for some forcing notion in $V$, and suppose the forcing extension satisfies the statement. Written out, this means \begin{align*} V[G]\models \exists x\in \omega^\omega\,\forall y\in \omega^\omega\, P(x,y,a). \end{align*} Since $a\in V$ is a ground-model real parameter and the displayed assertion is $\Sigma^1_2(a)$, Shoenfield absoluteness above applies and gives \begin{align*} V\models \exists x\in \omega^\omega\,\forall y\in \omega^\omega\, P(x,y,a). \end{align*} Thus forcing cannot make this $\Sigma^1_2(a)$ existence statement true for the first time. The particular witness $x$ seen in $V[G]$ may be a new real, but the ground model already has some witness to the same projective assertion. [/example] ## Martin-Steel Projective Determinacy How do Woodin cardinals influence concrete definable sets of reals? Their central descriptive-set-theoretic consequence is projective determinacy. The theorem is not proved in this course, since the proof uses inner model theory and iteration trees, but it is stated as the main bridge from large-cardinal strength to regularity for projective sets. [definition: Projective Determinacy] Projective determinacy is the assertion that every projective subset of $\omega^\omega$ is determined. [/definition] Projective determinacy is a scheme, one assertion for each projective level. The large-cardinal theorem calibrates this scheme by placing Woodin cardinals exactly where the projective hierarchy requires them. [quotetheorem:7444] The result is quoted here as the course's bridge theorem between large cardinals and definable games. Its hypotheses should be read level by level: $m$ Woodin cardinals below a measurable cardinal give control through the $(m+1)$st projective level, not through all projective levels at once. The measurable-above clause supplies the comparison and closure needed in the proof to build iterable extender models whose strategies decide the relevant games. The scope of the theorem is also sharp in two concrete ways. First, it gives determinacy for projective payoff sets, not for arbitrary subsets of $\omega^\omega$; under the full axiom of choice there are subsets of $\omega^\omega$ coding undetermined games. Second, below the stated large-cardinal strength, higher projective determinacy can fail in canonical inner models. For example, the constructible universe $L$ has no Woodin cardinals and contains projective well-orderings of the reals; these well-orderings are incompatible with the regularity consequences of projective determinacy, such as every uncountable projective set containing a perfect subset. Thus the Woodin-cardinal hypotheses are not decorative background assumptions: they mark the point at which the projective hierarchy stops behaving like the hierarchy inside $L$ and starts behaving like the Borel hierarchy with respect to games and regularity. [remark: Strength of the Hypothesis] For each fixed projective level there is a finite large-cardinal hypothesis sufficient for determinacy at that level: $m$ Woodin cardinals below a measurable cardinal suffice for $\Pi^1_{m+1}$ and $\Sigma^1_{m+1}$ determinacy. The final sentence of the theorem packages these finite implications into the projective-determinacy scheme. [/remark] The content of the theorem is not merely that games are decided. Once determinacy is available, the structure theory of definable sets becomes much closer to the Borel theory. [example: A Projective Game from Woodin Cardinals] Let $A\subseteq\omega^\omega$ be defined by a $\Pi^1_3$ formula. Concretely, after coding tuples of reals as reals, we may write membership in the form \begin{align*} z\in A \iff \forall u\in\omega^\omega\ \exists v\in\omega^\omega\ \forall n\in\mathbb N\ R(z,u,v,n), \end{align*} where $R$ is arithmetical. The payoff game $G_A$ is therefore the game in which Players I and II alternately play natural numbers, producing a real $z\in\omega^\omega$, and Player I wins exactly when the displayed $\Pi^1_3$ condition holds of $z$. Assume there are two Woodin cardinals $\delta_1<\delta_2$ and a measurable cardinal $\kappa>\delta_2$. Taking $m=2$ in the the [Martin-Steel projective determinacy theorem](/theorems/7444) above gives determinacy for every $\Pi^1_{2+1}$ subset of $\omega^\omega$. Since $2+1=3$, the set $A$ is determined. Thus exactly one of the following alternatives holds: either Player I has a strategy $\sigma$ such that every play following $\sigma$ produces $z\in A$, or Player II has a strategy $\tau$ such that every play following $\tau$ produces $z\notin A$. The winning strategy is not obtained by syntactically rewriting the formula defining $A$; the theorem asserts its existence from the large-cardinal hypothesis. This is the point of the example: cardinals high in the cumulative hierarchy decide a concrete game whose moves are only natural numbers. [/example] ## Regularity Properties from Determinacy What do we gain once every projective game is determined? The classical payoff is that projective sets satisfy the regularity properties familiar from Borel and analytic sets. Determinacy supplies uniform strategies, and those strategies rule out the pathological decompositions made possible by arbitrary choice. [definition: Regularity Properties for Sets of Reals] Let $A\subseteq \mathbb R$. 1. $A$ has the perfect set property if either $A$ is countable or $A$ contains a nonempty perfect subset. 2. $A$ has the Baire property if there is an open set $U\subseteq \mathbb R$ such that $A\triangle U$ is meagre. 3. $A$ is Lebesgue measurable if it is measurable with respect to the completion of Lebesgue measure. [/definition] These three properties express, respectively, size regularity, category regularity, and measure regularity. Since arbitrary sets of reals can fail them under choice, the natural question is whether determinacy for a definable pointclass is strong enough to recover all three properties inside that pointclass. [quotetheorem:7445] [citeproof:7445] The theorem explains why projective determinacy is visible far below the cardinals used to prove it, but the hypothesis cannot simply be discarded at higher projective levels. Under the axiom of choice there are arbitrary sets of reals without the Baire property and non-Lebesgue-measurable sets, and there are also undetermined games; projective determinacy says that the projective sets avoid these pathologies. The conclusion is regularity, not simplicity: a projective set with the Baire property or Lebesgue measurability need not be Borel, and the perfect set property does not classify the set beyond ruling out intermediate-size definable counterexamples. Thus the theorem gives strong structural control while still leaving the projective hierarchy genuinely richer than the Borel hierarchy. [example: Why Analytic Sets Are Well Behaved] Analytic sets already have the perfect set property, the Baire property, and Lebesgue measurability in ZFC. To see the mechanism behind these regularity facts, write an analytic set $A\subseteq\omega^\omega$ as the projection of a closed set $C\subseteq\omega^\omega\times\omega^\omega$, so \begin{align*} x\in A \iff \exists y\in\omega^\omega\ (x,y)\in C. \end{align*} For each $x\in\omega^\omega$, define the section tree \begin{align*} T_x=\{s\in\omega^{<\omega}:\exists y\in\omega^\omega\ (s\subseteq y\text{ and }(x,y)\in C)\}. \end{align*} If $x\in A$, choose $y$ with $(x,y)\in C$. Then for every $n\in\mathbb N$, the finite sequence $y|n$ lies in $T_x$, so $T_x$ has the infinite branch $y$. Conversely, suppose $T_x$ has an infinite branch $y$, meaning $y|n\in T_x$ for every $n$. For each $n$, the definition of $T_x$ gives some $y_n$ such that $y|n\subseteq y_n$ and $(x,y_n)\in C$. Thus every basic neighborhood determined by $y|n$ meets the closed section $C_x=\{u:(x,u)\in C\}$. Hence $y$ lies in the closure of $C_x$, and since $C_x$ is closed, $y\in C_x$. Therefore \begin{align*} x\in A \iff T_x\text{ has an infinite branch}. \end{align*} When $T_x$ is well-founded, assign ranks from the leaves upward by \begin{align*} \rho_x(s)=\sup\{\rho_x(s^\frown\langle k\rangle)+1:k\in\omega\text{ and }s^\frown\langle k\rangle\in T_x\}. \end{align*} Leaves have rank $0$ because the displayed supremum is over the empty set. Thus every $x$ is analyzed in one of two ways: either $T_x$ is ill-founded and a branch gives a witness $y$ to $x\in A$, or $T_x$ is well-founded and the ordinal rank records exactly how the search for such a witness fails. This rank-versus-branch dichotomy is the ZFC ancestor of the scale arguments used under projective determinacy. [/example] ## Generic Absoluteness from Woodin Cardinals Can large cardinals make higher projective truth as stable under forcing as Shoenfield truth is at level $\Sigma^1_2$? Woodin cardinals were designed in part to answer this kind of question. Their generic absoluteness consequences say that forcing cannot alter substantial fragments of the theory of the reals. [definition: Projective Generic Absoluteness] A projective statement $\varphi(a)$ with real parameter $a$ is generically absolute over a model $V$ if for every forcing notion $\mathbb P\in V$ and every $V$-generic filter $G\subseteq \mathbb P$, one has \begin{align*} V\models \varphi(a) \iff V[G]\models \varphi(a). \end{align*} [/definition] This definition extends Shoenfield's conclusion to higher complexity. The difference is that now the proof requires the large-cardinal machinery developed earlier in the course. Woodin-cardinal generic absoluteness is used here as a landmark rather than as a theorem proved locally. At the base level, Shoenfield absoluteness already gives forcing invariance for $\Sigma^1_2$ and $\Pi^1_2$ statements with ground-model real parameters. Finite stacks of Woodin cardinals, usually with a measurable cardinal above in the standard calibration, push this invariance upward through finite projective levels. The statement is level-by-level because the large-cardinal strength is part of the content: the jump from $\Sigma^1_2$ to higher projective complexity requires new Woodin-cardinal strength. The restriction to parameters from the ground model is essential; if forcing adds a new real $b$, then statements using $b$ may express information that was not present in $V$ and are not covered by this kind of absoluteness. There is a concrete forcing-sensitive boundary behind the hypothesis. In $L$, the canonical projective well-ordering of the reals can be used to build higher projective statements whose truth changes after adding Cohen or collapse generics, because the extension has new reals while the old projective definition still refers to the ground-model constructibility hierarchy. More generally, without Woodin-cardinal generic absoluteness, projective assertions beyond the Shoenfield level can distinguish the ground model from forcing extensions. Woodin's stationary tower, extender algebra, and iteration-strategy machinery prevent this by making the relevant finite fragment of projective truth invariant under all set forcing. [remark: Relation to Shoenfield Absoluteness] Shoenfield absoluteness is the zero-large-cardinal case of the same pattern: low-complexity projective truth is already stable. Woodin cardinals push this stability upward through the projective hierarchy, matching the determinacy consequences supplied by Martin-Steel. [/remark] This final perspective ties together the chapter. Projective determinacy gives regularity and scales, while generic absoluteness gives forcing invariance. Both say that, under Woodin-cardinal hypotheses, definable sets of reals form a stable mathematical universe rather than a forcing-sensitive collection of pathologies. Descriptive set theory reveals one major use of Woodin-cardinal strength: it stabilizes definable sets of reals against forcing and pathology. Chapter 11 abstracts this lesson into calibration, treating large cardinals as a scale for measuring the exact strength of mathematical principles beyond ZFC. # 11. Consistency Strength and Calibration The guiding theme is calibration. Large cardinal theory is not a list of unrelated higher infinities: it is a scale used to measure the strength of mathematical principles that cannot be settled in ZFC alone. The chapter assumes the earlier material on the cumulative hierarchy, inaccessible, weakly compact, measurable, strong, and Woodin cardinals, together with the basic metamathematics of first-order theories and Gödel incompleteness. The scale is subtle, because many comparisons are statements about consistency rather than implications inside ZFC. ## Relative Consistency and Equiconsistency How can we compare two theories when ZFC cannot prove that either theory has a model? The method is to compare what would follow from the existence of a model of one theory. This is weaker than proving the theory outright, but it is the right level of comparison once Gödel's incompleteness theorems enter the picture. [definition: Relative Consistency] Let $T$ and $S$ be first-order theories. We say that $T$ is relatively consistent with $S$ if there is a proof, in a fixed weak metatheory, of \begin{align*} \operatorname{Con}(S) \implies \operatorname{Con}(T). \end{align*} [/definition] The phrase records a metamathematical implication between consistency statements. It does not say that $S$ proves $T$, and it does not say that every model of $S$ satisfies $T$. To compare theories symmetrically, we need the case where each theory can play the role of the stronger consistency assumption for the other. [definition: Equiconsistency] Let $T$ and $S$ be first-order theories. We say that $T$ and $S$ are equiconsistent if both relative consistency implications hold: \begin{align*} \operatorname{Con}(T) \implies \operatorname{Con}(S). \end{align*} \begin{align*} \operatorname{Con}(S) \implies \operatorname{Con}(T). \end{align*} [/definition] Equiconsistency is a comparison of consistency strength, not a comparison of deductive content. Two theories can be equiconsistent while proving very different mathematical statements, because the comparison ignores everything except whether a contradiction can be derived. [example: Inaccessible Versus ZFC] Assume that $\kappa$ is a strongly inaccessible cardinal. By the rank-initial model theorem above, the rank-initial segment $V_\kappa$ satisfies every axiom of ZFC, so \begin{align*} V_\kappa \models \mathrm{ZFC}. \end{align*} In the metatheory, this gives a concrete model of ZFC from the stronger hypothesis $\mathrm{ZFC}+\text{``there is an inaccessible cardinal''}$. Therefore, if the stronger theory has a model containing such a $\kappa$, then its internal $V_\kappa$ is a model of ZFC, and hence \begin{align*} \operatorname{Con}(\mathrm{ZFC}+\text{``there is an inaccessible cardinal''}) \implies \operatorname{Con}(\mathrm{ZFC}). \end{align*} This places the inaccessible-cardinal axiom above bare ZFC in consistency strength: the argument does not make ZFC prove its own consistency, but uses the stronger inaccessible-cardinal hypothesis to build a model of ZFC. [/example] The preceding example also explains why consistency strength is directional. To use this comparison systematically, we need the formal model-building fact behind it: strong inaccessibility gives exactly enough closure for a rank-initial segment to satisfy the ZFC axioms. [quotetheorem:7446] [citeproof:7446] The hypotheses are doing real work. If $\kappa$ is not regular, Replacement can fail in $V_\kappa$: a cofinal sequence in $\kappa$ may be the range of a function whose domain has rank below $\kappa$ but whose range is unbounded in rank. If $\kappa$ is not a limit stage with the needed rank closure, Power Set can fail for sets whose subsets first appear above the stage. The theorem also does not say that ZFC proves there is such a $\kappa$, nor that every model of ZFC is of the form $V_\kappa$. Its role is forward-looking: many later consistency comparisons use stronger cardinals to build not only rank-initial models, but also ultrapowers, extender models, and inner models. ## A Working Order of Large Cardinal Strength Which cardinals should be placed above which on the large cardinal scale? Some implications hold directly in ZFC: a measurable cardinal is weakly compact, and a weakly compact cardinal is inaccessible. Other comparisons are consistency-strength comparisons: a Woodin cardinal does not merely sit above a single local embedding property, but organizes many such properties across smaller ranks. [definition: Consistency Strength Ordering] For theories $T$ and $S$ extending ZFC, write $T \leq_{\operatorname{Con}} S$ if \begin{align*} \operatorname{Con}(S) \implies \operatorname{Con}(T) \end{align*} is provable in the chosen metatheory. [/definition] This relation is a preorder rather than a linear order. It is reflexive and transitive, but two theories can be incomparable by current methods, and two distinct theories can be equiconsistent. [example: A Consistency-Strength Chart] Let \begin{align*} T_0=\mathrm{ZFC} \end{align*} and let $T_1,T_2,T_3,T_4$ be ZFC plus, respectively, “there is an inaccessible cardinal,” “there is a Mahlo cardinal,” “there is a weakly compact cardinal,” and “there is a measurable cardinal.” The standard first part of the consistency-strength chart is \begin{align*} T_0\leq_{\operatorname{Con}}T_1\leq_{\operatorname{Con}}T_2\leq_{\operatorname{Con}}T_3\leq_{\operatorname{Con}}T_4. \end{align*} Here $T_i\leq_{\operatorname{Con}}T_{i+1}$ means exactly that \begin{align*} \operatorname{Con}(T_{i+1})\implies \operatorname{Con}(T_i). \end{align*} The first step is witnessed by the rank-initial model theorem above: from a model of $T_1$ with inaccessible $\kappa$, the internal rank segment $V_\kappa$ is a model of $\mathrm{ZFC}$, so $\operatorname{Con}(T_1)\implies\operatorname{Con}(T_0)$. The next steps come from the corresponding implication of large-cardinal properties: a Mahlo cardinal is inaccessible, a weakly compact cardinal is Mahlo, and a measurable cardinal is weakly compact by the theorem above that [measurable cardinals are weakly compact](/theorems/7447). Thus a model with the stronger cardinal also contains the weaker kind of cardinal required for the preceding theory. Stronger embedding hypotheses continue the calibration, but the chart should not be read as one simple linear list of one-cardinal implications. A strong cardinal is a property of one critical point relative to rank targets; a Woodin cardinal gives many local strongness witnesses below itself, organized by functions $f:\delta\to\delta$; and supercompactness imposes still broader closure requirements across many sizes. Therefore the usable comparison is not the slogan “strong $<$ Woodin $<$ supercompact,” but the precise statement being compared, such as \begin{align*} \operatorname{Con}(\mathrm{ZFC}+\text{``there is a measurable cardinal''})\implies \operatorname{Con}(\mathrm{ZFC}+\text{``there is a weakly compact cardinal''}). \end{align*} The chart is a map of standard implication and relative-consistency comparisons, not a proof that every natural large-cardinal axiom lies on a single line; its purpose is to locate which hypotheses are strong enough for particular independence, regularity, and determinacy arguments. [/example] The chart becomes mathematically meaningful only because some adjacent steps are witnessed by theorems rather than by convention. The first serious obstruction is that weak compactness is usually recognized through reflection or the tree property, while measurability is phrased through a $\kappa$-complete ultrafilter. A direct combinatorial construction of branches through arbitrary $\kappa$-trees is hard because there is no canonical choice of compatible nodes at all levels. The ultrapower mechanism supplies exactly the missing global object, so the next theorem gives an embedding-based comparison: measurability entails weak compactness. [quotetheorem:7447] [citeproof:7447] The hypothesis is stronger than the conclusion. There are models, relative to standard large-cardinal assumptions, in which weakly compact cardinals exist but no measurable cardinal exists; thus the theorem is not reversible. It also does not say that every proof of weak compactness must use an ultrafilter, only that measurability provides one efficient route to the tree property. This distinction prepares the next stage of calibration: embedding hypotheses often imply reflection or compactness properties, but the reverse direction usually requires extra assumptions or is false in consistency strength. [remark: Implication Versus Consistency Strength] If ZFC proves that axiom $A$ implies axiom $B$, then \begin{align*} \operatorname{Con}(\mathrm{ZFC}+A) \implies \operatorname{Con}(\mathrm{ZFC}+B). \end{align*} The converse need not hold. Equiconsistency can exist without any known implication between the corresponding large cardinal properties. [/remark] This distinction matters when comparing strong, Woodin, and supercompact cardinals. Many standard implications are known, but the hierarchy is refined by consistency-strength results involving inner models, forcing extensions, and descriptive set-theoretic consequences. ## Reflection, Embeddings, and Determinacy Why do the same cardinals appear in arguments that look so different? The course has used three languages for large cardinal strength: reflection principles, elementary embeddings, and determinacy principles. Calibration often means translating among these languages without pretending they are identical. [definition: Reflection Language] A large cardinal axiom is expressed in the reflection language when it asserts that structures, formulas, stationary sets, or combinatorial configurations visible at a high rank already occur correctly below that rank. [/definition] Reflection is the natural language for inaccessible, Mahlo, and weakly compact cardinals. It emphasizes the universe looking internally coherent at some stage $V_\kappa$ and often produces compactness phenomena. [example: Weak Compactness as Reflection] Let $\kappa$ be inaccessible, and let $A\subset V_\kappa$. The reflection formulation says that $\kappa$ is weakly compact exactly when every $\Pi^1_1$ statement true in the structure $(V_\kappa,\in,A)$ is already true in some smaller initial segment with the parameter restricted to that segment. Concretely, if \begin{align*} (V_\kappa,\in,A)\models \varphi \end{align*} for a $\Pi^1_1$ sentence $\varphi$, then there is some $\alpha<\kappa$ such that \begin{align*} (V_\alpha,\in,A\cap V_\alpha)\models \varphi. \end{align*} The parameter is changed from $A$ to $A\cap V_\alpha$ because the smaller structure only sees sets of rank below $\alpha$. Thus weak compactness can be read as a precise reflection principle: any $\Pi^1_1$ property visible at the inaccessible level $V_\kappa$ has a correct smaller witness $V_\alpha$. This is the reflection-language version of the same large-cardinal strength that the tree-property formulation measures combinatorially. [/example] Reflection explains the lower part of the hierarchy, but it does not by itself display the mechanisms behind measurable, strong, Woodin, and supercompact cardinals. The next definition introduces the embedding language, which is needed because it records how much of the universe can be moved while preserving first-order truth. [definition: Embedding Language] A large cardinal axiom is expressed in the embedding language when it asserts the existence of a non-identity elementary embedding $j:M\to N$ between models of set theory, usually with critical point $\kappa$ and additional closure or target requirements. [/definition] The critical point measures where the embedding first moves the universe. Stronger axioms require the embedding to see more of the universe correctly, to have more closure in the target model, or to exist for larger classes of parameters. In the function formulation of Woodinness, the data are organized as follows: for every function $f:\delta\to\delta$, some critical point $\kappa<\delta$ is a closure point of $f$, and an elementary embedding with critical point $\kappa$ sees the rank initial segment $V_{j(f)(\kappa)}$. The closure-point hypothesis is essential: without $f[\kappa]\subseteq\kappa$, the function could place the required bound above the proposed critical point before the embedding has any chance to witness the intended local strength. A concrete boundary case is the following. Fix a candidate critical point $\kappa<\delta$ and define $f(\alpha)=\kappa+1$ for all $\alpha<\kappa$; then $f[\kappa]\nsubseteq\kappa$, so that same $\kappa$ is disqualified before any embedding conclusion is considered. This prevents the definition from counting a critical point whose earlier parameter values already outrun it. The rank-inclusion hypothesis is also necessary: an elementary embedding with critical point $\kappa$ alone, with no requirement that $V_{j(f)(\kappa)}\subset M$, would not say that $\kappa$ has any prescribed local strongness. The theorem also does not assert that $\delta$ itself is strong, nor that a Woodin cardinal is just a stronger version of a single strong cardinal. Rather, it says that below $\delta$ there are many critical points adapted to arbitrary functions $f:\delta\to\delta$. This coordinated supply of embeddings is the feature that later links Woodin cardinals to determinacy and to inner model comparison arguments. [definition: Determinacy Language] A principle is expressed in the determinacy language when it asserts that certain infinite two-player perfect-information games are determined, meaning that one of the players has a winning strategy. [/definition] Determinacy statements are not large cardinal axioms syntactically, but they often have large-cardinal consistency strength. The calibration problem is to determine whether projective regularity principles are merely consequences of special game arguments or whether they correspond to precise embedding strength. The Martin-Steel analysis places finite projective determinacy levels on the same consistency-strength scale as finite stacks of Woodin cardinals, with a measurable cardinal above in the standard formulation used earlier. The quantification over finite levels is necessary. A fixed finite number of Woodin cardinals calibrates only a fixed finite projective fragment, so it does not by itself give full projective determinacy. The indexing is also part of the content: one Woodin cardinal with a measurable above is associated with low projective determinacy in this calibration, while stronger finite fragments require additional Woodin cardinals. A boundary case is supplied by the finite mouse hierarchy: the $n$-Woodin level gives the lower bound for the corresponding finite projective determinacy level, but it is not the calibration for the next level; moving up one projective level raises the required mouse strength. This is not a new theorem proved here, but a summary of how the quoted Martin-Steel results are used to locate regularity properties of definable sets of reals on the same consistency-strength map as embedding axioms. ## What ZFC Alone Cannot Prove What does calibration tell us about ordinary mathematical statements? It tells us that many natural principles are neither arbitrary nor free-floating: they occupy specific regions of the consistency-strength hierarchy. ZFC may leave them undecided, while large cardinal hypotheses explain why they are stable, useful, and difficult to refute by ordinary means. Gödel's second incompleteness theorem sets the background limitation. If ZFC is consistent, then ZFC does not prove its own standard arithmetized consistency statement. More generally, if $T_A=\mathrm{ZFC}+A$ is a recursively axiomatized extension of ZFC, then a ZFC proof of $\operatorname{Con}(T_A)$ would also give a ZFC proof of $\operatorname{Con}(\mathrm{ZFC})$, since $T_A$ extends ZFC. Thus, under the usual consistency assumption on ZFC and the standard proof-coding formalization, ZFC cannot certify the consistency of such stronger recursively axiomatized large-cardinal theories. The hypotheses of Gödel's theorem matter here. The theory under discussion must be recursively axiomatized and strong enough to formalize the relevant syntax and proof predicate; without those coding assumptions, the theorem is not the same statement. A concrete boundary case is obtained by adding all true first-order sentences of arithmetic to a weak arithmetical base: the resulting complete theory can assert its own consistency in the intended interpretation, but it is not recursively axiomatized, so Gödel's second incompleteness theorem does not apply in the same form. Another boundary case is a theory too weak to represent the proof predicate needed for the diagonal argument. The result also does not say that $\mathrm{ZFC}+A$ is inconsistent, only that ZFC cannot certify its consistency if ZFC itself is consistent and the formalization is the usual one. This is why large cardinal justification is not expected to be a ZFC proof of consistency. Instead, the next kind of evidence comes from coherence, reflection, consequences, inner model theory, forcing calibrations, and the position of the axiom in the wider hierarchy. [example: Comparing Measurable and Inaccessible Consistency] Suppose $\kappa$ is measurable, and let $U$ be a normal measure on $\kappa$. By the theorem above that measurable cardinals are weakly compact, $\kappa$ is weakly compact; in particular, $\kappa$ is strongly inaccessible. Therefore the rank-initial model theorem above applies to $\kappa$, so \begin{align*} V_\kappa \models \mathrm{ZFC}. \end{align*} It remains to see that $V_\kappa$ contains an inaccessible cardinal. Let \begin{align*} X=\{\alpha<\kappa:\alpha \text{ is strongly inaccessible}\}. \end{align*} For a normal measure $U$, the standard ultrapower criterion says \begin{align*} X\in U \quad \text{if and only if} \quad \kappa\in j(X), \end{align*} where $j:V\to M$ is the ultrapower embedding by $U$. Since $\operatorname{crit}(j)=\kappa$, the embedding fixes every ordinal below $\kappa$, and $j(X)$ is the corresponding set of inaccessible cardinals below $j(\kappa)$. The cardinal $\kappa$ is inaccessible in $M$, so \begin{align*} \kappa\in j(X). \end{align*} Hence \begin{align*} X\in U. \end{align*} Because $U$ is a non-principal ultrafilter on $\kappa$, no empty set belongs to $U$, so $X\neq\varnothing$. Choose $\lambda\in X$. Then \begin{align*} \lambda<\kappa \end{align*} and $\lambda$ is strongly inaccessible. Since $\lambda<\kappa$, the relevant ranks witnessing the inaccessibility of $\lambda$ are already present in $V_\kappa$, so \begin{align*} V_\kappa \models \text{``there is an inaccessible cardinal''}. \end{align*} Thus a model of $\mathrm{ZFC}+\text{``there is a measurable cardinal''}$ yields a model of $\mathrm{ZFC}+\text{``there is an inaccessible cardinal''}$, giving \begin{align*} \operatorname{Con}(\mathrm{ZFC}+\text{``there is a measurable cardinal''})\implies \operatorname{Con}(\mathrm{ZFC}+\text{``there is an inaccessible cardinal''}). \end{align*} The comparison places measurability above inaccessibility in consistency strength, while still making only a metamathematical consistency comparison rather than a ZFC proof of strictness. [/example] The final perspective is that large cardinal axioms function as coordinates. When a theorem requires a measurable cardinal, a Woodin cardinal, or a supercompact cardinal, the hypothesis is not merely a technical assumption; it locates the theorem on the map of consistency strength. [remark: Calibration as a Course Summary] Inaccessible and Mahlo cardinals express closure and reflection inside the cumulative hierarchy. Weakly compact cardinals add higher-order reflection and compactness. Measurable cardinals introduce ultrapower embeddings. Strong and Woodin cardinals organize embeddings across larger rank segments, while supercompact cardinals impose sweeping closure across many sizes. Determinacy translates part of this hierarchy into the language of games on the reals. [/remark] The course began with $V_\kappa$ as a model-building device and ends with large cardinals as a measuring system for mathematics beyond ZFC. The same pattern recurs throughout modern set theory: formulate a natural principle, locate its consistency strength, and then use that location to understand both its consequences and its limitations. Calibration turns the hierarchy into a tool for comparing theories, consequences, and limitations. The capstone chapter now gathers the course's themes into the modern picture, where large cardinals, inner models, HOD, and generic absoluteness interact as parts of one research program. # 12. Capstone: The Modern Picture This capstone chapter gathers the main themes of the course into the modern research picture around large cardinals, canonical inner models, HOD, and generic absoluteness. The preceding chapters developed the technical language of reflection, ultrapowers, elementary embeddings, extenders, determinacy strength, and forcing; here those tools are used as a map for reading contemporary statements rather than as isolated definitions. The guiding question is how large-cardinal axioms organise the universe of sets: do they point toward canonical inner models, stable theories under forcing, or a more plural landscape of possible universes? ## Ultimate $L$ and the Search for Canonical Universes What would it mean for the universe of sets to have a canonical core after the discovery of forcing? Gödel's $L$ gives a rigid and definable universe, but it is too small for much of the large-cardinal hierarchy. The contemporary inner model program asks for canonical models that retain the fine structure and explanatory power of $L$ while accommodating stronger large cardinals. [definition: Canonical Inner Model Program] The canonical inner model program is the project of constructing definable proper class inner models $M$ of the ambient universe $V$ satisfying ZFC and carrying prescribed large-cardinal structure, with fine structural analysis sufficient to compare such models by iteration. [/definition] The phrase "canonical" is doing substantial work here. It asks that the model be definable without arbitrary choices, sufficiently close to $V$ to analyse sets and cardinals, and sufficiently rigid that comparison arguments determine it up to a meaningful equivalence. [example: From $L$ To Core Models] The constructible universe $L$ is the prototype of a canonical inner model: it is defined stage by stage by \begin{align*} L_0=\varnothing,\qquad L_{\alpha+1}=\operatorname{Def}(L_\alpha),\qquad L_\lambda=\bigcup_{\alpha<\lambda}L_\alpha \end{align*} for limit $\lambda$, and \begin{align*} L=\bigcup_{\alpha\in\operatorname{Ord}}L_\alpha . \end{align*} This definition uses only the ordinals and first-order definability, so $L$ is not built from arbitrary choices of parameters outside the construction. Its fine structure then supports condensation, covering, and comparison arguments, which is why it functions as the basic model of canonicity. The limitation appears as soon as the intended universe contains measurable cardinals. A measurable cardinal is not represented by a literal set-sized elementary embedding $j:V\to M$ inside $V$; instead, it is represented by a normal measure $U$ on $\kappa$, from which one forms the ultrapower map \begin{align*} j_U:V\to \operatorname{Ult}(V,U). \end{align*} The embedding data is therefore encoded by the measure $U$, and stronger large cardinals require longer coherent systems of such measure-like objects, namely extenders. Since the construction of $L$ only adds sets definable over earlier levels and does not add canonical extender sequences, plain $L$ is too small to carry the large-cardinal structure that these hypotheses demand. Core models such as $K$ keep the same guiding principle as $L$ but alter the data being constructed: below a specified large-cardinal threshold, the model is built with exactly the coherent extender information allowed by that threshold. Thus $L$ is the canonical model when no extender structure is being used, while $K$ is the corresponding canonical substitute when the relevant measurable or extender-like structure must be recorded. The recurring pattern is that, when a large cardinal is absent, a canonical inner model can often replace the ambient universe for comparison arguments and yield strong combinatorial consequences. [/example] This example explains why $L$ remains the model for canonicity, but it also exposes the obstruction: the original construction cannot house the upper part of the hierarchy studied in this course. The next problem is therefore to name the hoped-for analogue of $L$ that would retain canonicity while allowing very strong large cardinals. [definition: Ultimate $L$] Ultimate $L$ is an informal name for a proposed canonical inner model extending the role of $L$ to a setting compatible with large cardinals far beyond measurability, especially supercompact-like strength. [/definition] This definition is deliberately informal because the full program is not a single elementary construction in the way that $L$ is. In these notes, Ultimate $L$ should be read as a research aim: a possible canonical universe whose existence would organise the relationship between large cardinals, HOD, forcing axioms, and generic absoluteness. [remark: Why Ultimate $L$ Is Not Just $L$ With Large Cardinals] The original $L$ satisfies strong combinatorial principles such as $V=L$, but it rules out many large cardinals. A successful Ultimate $L$ program would have to preserve the advantages of constructibility while avoiding the collapse of the higher large-cardinal hierarchy. This tension is one reason the program is central rather than cosmetic. [/remark] The preceding remark gives a test for any proposed canonical universe: it must explain not only which sets it contains, but also how much of $V$ is recoverable from definable information. That question leads directly to HOD, the inner model that records hereditary definability from ordinals. [definition: HOD] The class $\text{HOD}$ consists of all sets $x$ such that every element of the transitive closure of $\{x\}$ is ordinal-definable in $V$. [/definition] HOD is not introduced here as a technical replacement for $L$, but as a diagnostic object. Once HOD is available, the natural large-cardinal question is whether this definable inner model resembles $V$ enough to inherit large-cardinal strength or whether it diverges from $V$ in a systematic way. [explanation: HOD Dichotomy As A Research Guidepost] Modern HOD dichotomy results ask whether, under strong large-cardinal hypotheses, HOD behaves like a close canonical approximation to $V$ or instead fails to approximate the large-cardinal structure of $V$ in a systematic way. In the favourable case, HOD may contain substantial large-cardinal structure of its own; in the opposite case, covering and approximation failures witness that hereditary ordinal definability is too sparse to reflect the ambient universe. This is not a single theorem at the level of these notes. Precise versions require specifying the large-cardinal assumptions, the exact covering property, and the internal large-cardinal structure being tested. Their proofs use extender models, HOD analysis under determinacy, and comparison theory that belongs to a later course. [/explanation] The point of stating the dichotomy in this schematic form is to train the reader to parse research statements without confusing the slogan with a fixed theorem. The next example translates a typical dichotomy-style sentence into the vocabulary developed in the course. [example: Reading The HOD Dichotomy] Suppose a paper states: If there is a proper class of Woodin cardinals, then either $\mathrm{HOD}$ has a proper class of measurable cardinals, or $\mathrm{HOD}$ is far from $V$ in a precise covering sense. The hypothesis has three separate pieces to read. First, "a proper class of Woodin cardinals" means that for every ordinal $\alpha$ there is a Woodin cardinal $\delta>\alpha$; the assumption is therefore not about one isolated large cardinal, but about unbounded Woodin strength throughout the universe. Second, $\mathrm{HOD}$ is the inner model being tested: its sets are those whose transitive closures are built from objects ordinal-definable in $V$. Third, the first alternative says that $\mathrm{HOD}$ itself contains arbitrarily large cardinals that it regards as measurable. The dichotomy can then be parsed as follows. In the first case, definability from ordinals has preserved substantial large-cardinal structure: for every ordinal $\alpha$, there is a $\kappa>\alpha$ such that $\mathrm{HOD}$ has a normal measure on $\kappa$ and so regards $\kappa$ as measurable. In the second case, $\mathrm{HOD}$ does not approximate $V$ well with respect to the relevant covering property: sets or cardinals in $V$ cannot be uniformly captured by small-enough sets from $\mathrm{HOD}$ in the required way. Thus the conclusion is not merely that $\mathrm{HOD}$ is different from $V$; it says that HOD analysis must fall into one of two structural regimes, either inheriting large-cardinal strength or failing a concrete approximation test. [/example] ## Large Cardinals as Axioms for Structural Stability Why should axioms asserting enormous infinities affect ordinary mathematical structure? The answer developed through the course is that large cardinals often produce absoluteness, reflection, compactness, and comparison principles. These are stability principles: they say that truths do not change arbitrarily when viewed through smaller ranks, inner models, ultrapowers, forcing extensions, or definable fragments of the universe. [definition: Structural Stability Principle] A structural stability principle is a set-theoretic assertion stating that a class of mathematical truths, constructions, or comparisons is invariant under a specified family of transformations such as rank reflection, elementary embeddings, forcing extensions, or passage to canonical inner models. [/definition] This terminology is not a new axiom scheme. It is a way of organising the role played by the large cardinals we have studied: each level of the hierarchy grants a stronger form of stability. [example: Stability Across The Hierarchy] At the first level, let $\kappa$ be strongly inaccessible. The closure and regularity conditions on $\kappa$ ensure that the usual set-forming operations needed for ZFC stay below rank $\kappa$: if $x\in V_\kappa$, then $\operatorname{rank}(x)<\kappa$, so pairing, unions, subsets, and replacement images have ranks bounded below $\kappa$. Thus $V_\kappa$ is not merely a large set; it is an initial segment in which the ZFC axioms can be checked internally. At the next level, let $\kappa$ be measurable and let $U$ be a normal measure on $\kappa$. The ultrapower construction forms equivalence classes $[f]_U$ of functions $f:\kappa\to V$, where $f\sim_U g$ means $\{\alpha<\kappa:f(\alpha)=g(\alpha)\}\in U$. The associated map is \begin{align*} j_U(x)=[c_x]_U, \end{align*} where $c_x(\alpha)=x$ for every $\alpha<\kappa$. Łoś's theorem for ultrapowers says that a first-order formula holds of the classes $[f_1]_U,\ldots,[f_n]_U$ exactly when the set of $\alpha<\kappa$ where it holds of $f_1(\alpha),\ldots,f_n(\alpha)$ lies in $U$. Hence first-order truth is transported coherently from $V$ to $\operatorname{Ult}(V,U)$. Woodin cardinals push the same stability pattern into the theory of reals. Instead of only producing one ultrapower, they give enough embedding and extender structure to control definable sets of reals, especially projective and universally Baire sets. In that setting, the stable object is not an initial segment like $V_\kappa$ or one ultrapower target, but the truth of statements about sets of reals across forcing extensions. Thus inaccessible, measurable, and Woodin cardinals exhibit the same organizing theme at increasing strength: closure stabilizes axioms, ultrapowers stabilize first-order truth, and Woodin strength stabilizes definable real analysis under forcing. [/example] The hierarchy-wide example shows the same pattern at several strengths, so the next step is to revisit the first precise case where the pattern becomes a theorem. Inaccessibility is the entry point because its closure and regularity properties are exactly what is needed for an initial segment to satisfy all ZFC axioms. [quotetheorem:7448] [citeproof:7448] This proof is the template for many later arguments, but the hypotheses are not interchangeable. If $\kappa$ is singular, Replacement can fail in $V_\kappa$ because the ranks of a set-sized family of values may be cofinal in $\kappa$. If $\kappa$ is not a limit ordinal, the basic closure needed for the hierarchy breaks down, and if one asks for cardinal-size closure properties such as $|V_\alpha| < \kappa$ for all $\alpha < \kappa$, the strong limit condition is the relevant ingredient. The theorem also does not say that $V_\kappa$ is elementarily equivalent to $V$ or that it contains all subsets of its cardinals as computed in $V$; it says only that the internal first-order axioms of ZFC are satisfied. This is the forward pattern used throughout large-cardinal theory: extra height and closure turn an initial segment, inner model, or ultrapower into a universe-like object on which comparison and reflection arguments can run. [example: Research Hypotheses As Stability Clues] When a theorem assumes a supercompact cardinal and derives a forcing axiom, read the hypothesis as specifying the kind of stability the proof is allowed to use. Supercompactness says that for arbitrarily large targets $\lambda$, there are elementary embeddings $j:V\to M$ with critical point $\kappa$ and enough closure in the target model to keep $\lambda$-length information together, usually expressed as $M^\lambda\subseteq M$. In a forcing-axiom argument, this closure is what lets the construction handle many dense sets at once: if $\langle D_\xi:\xi<\lambda\rangle$ is a sequence of dense subsets of a forcing notion $\mathbb P$, then the relevant embedding-and-lifting argument is designed so that this whole sequence is visible in the target model, not merely each $D_\xi$ separately. The conclusion that a filter meets all the required dense sets is therefore a closure-and-coherence conclusion, not an accidental consequence of assuming a large cardinal. By contrast, when a theorem assumes a proper class of Woodin cardinals and proves generic absoluteness for a theory of $H_{\omega_1}$, the stability clue is different. “Proper class of Woodin cardinals” means that for every ordinal $\alpha$ there is a Woodin cardinal $\delta>\alpha$, so the hypothesis supplies Woodin strength unboundedly high in the universe. The conclusion says that, for the specified language and forcing class, if $G$ is generic, then the relevant statements about $H_{\omega_1}$ have the same truth value in $V$ and in $V[G]$. Thus supercompact hypotheses usually signal closure strong enough to build forcing extensions with coherent generic objects, while Woodin hypotheses usually signal control over definable sets of reals and preservation of truth across forcing extensions. [/example] This example separates closure-based stability from forcing-absoluteness stability, and it raises the next question: which large cardinals are designed to control forcing over the real line? This motivates the following informal theorem, because it records the specific stabilising role of Woodin cardinals. [explanation: Woodin Cardinals And Generic Absoluteness] Woodin cardinals are central to generic absoluteness because they give control over definable sets of reals, especially through the theory of universally Baire sets. In many precise theorems, a specified supply of Woodin cardinals implies that the truth of selected statements about structures built from the reals is preserved by a specified class of forcing extensions. The exact theorem depends on several parameters: the number and arrangement of Woodin cardinals, the language under consideration, the class of sets of reals allowed as parameters, and the forcing notions permitted. The message for this course is that Woodin cardinals are not only higher consistency-strength markers; they are also mechanisms for controlling the effect of forcing on definable mathematical truth. [/explanation] This forcing stability is closely related to the determinacy themes that appeared earlier, but the direction of emphasis changes. Generic absoluteness asks which truths survive forcing extensions; determinacy asks which sets of reals have regular structure, and large cardinals provide the bridge between those viewpoints. [remark: Determinacy As Another Stability Form] Under determinacy hypotheses, sets of reals acquire regularity properties such as the perfect set property, Lebesgue measurability, and the Baire property. Large cardinals give consistency strength for such determinacy theories, while determinacy supplies a stable descriptive set theory incompatible with full choice. This is one of the main reasons Woodin cardinals occupy a central position. [/remark] ## Open Directions in HOD, Supercompactness, and Generic Absoluteness Where does the course stop, and where does current research begin? The boundary is not at a single named cardinal. It is the point where the elementary-embedding methods of the course meet unresolved problems about canonical inner models, the exact structure of HOD, and the strongest possible forms of forcing absoluteness. The first direction is HOD analysis. HOD is definable in every universe of ZFC, but its structure can vary dramatically. Large cardinals sharpen the question by asking whether HOD inherits the large-cardinal strength of $V$ or instead reflects a deep gap between definability and the full universe. [definition: HOD Analysis] HOD analysis is the study of the internal structure of $\text{HOD}$, especially its cardinals, extenders, covering properties, and large-cardinal content, under ambient hypotheses on $V$. [/definition] The purpose of HOD analysis is not only to understand one inner model. It is also a way of testing whether proposed axioms make the universe more canonical or more multiverse-like. [example: HOD Under Determinacy] Under determinacy, HOD can be studied even when the ambient model does not satisfy full Choice. For instance, work in $L(\mathbb R)$ and assume $\operatorname{AD}$ there. The object being analysed is \begin{align*} \mathrm{HOD}^{L(\mathbb R)}=\{x\in L(\mathbb R):x\text{ and every member of its transitive closure is ordinal-definable in }L(\mathbb R)\}. \end{align*} Thus a set $x$ belongs to $\mathrm{HOD}^{L(\mathbb R)}$ exactly when there are ordinals $\alpha_1,\ldots,\alpha_n$ and a formula $\varphi$ such that $x$ is the unique object in $L(\mathbb R)$ satisfying $\varphi(x,\alpha_1,\ldots,\alpha_n)$, and the same ordinal-definability condition holds hereditarily for the elements appearing below $x$. A typical HOD-analysis theorem in this setting does not merely say that $\mathrm{HOD}^{L(\mathbb R)}$ is definable. It identifies internal structure: cardinals of $\mathrm{HOD}^{L(\mathbb R)}$, extender-like sequences, and comparison data that make HOD resemble a fine-structural inner model. The determinacy assumption supplies regularity and scale structure for sets of reals, while the large-cardinal background explains why such determinacy is consistent. The example therefore links three course themes in one place: inner models supply the object being analysed, ordinal definability supplies HOD, and Woodin-cardinal strength supplies the consistency-theoretic source of the determinacy hypothesis. [/example] The determinacy example shows that HOD analysis can reveal fine structure even in settings far from ordinary choice-based universes. The next open problem asks for a still stronger target: a canonical inner model whose own large-cardinal structure reaches supercompactness. [definition: Inner Model For Supercompactness] An inner model for supercompactness is a proper class inner model $M$ of the ambient universe $V$ such that $M$ satisfies ZFC and, for some cardinal $\kappa \in M$, the model $M$ satisfies that $\kappa$ is supercompact, with all measures, target models, and quantifiers interpreted internally to $M$. [/definition] The definition is short, but the construction problem is profound. Supercompactness involves a wide range of measures and embeddings, and canonical inner models must encode this information in a way that supports comparison. [remark: Why Supercompactness Is Hard For Core Models] Measurable and Woodin cardinals can be represented using extenders with enough coherence to support comparison theory. Supercompactness demands closure targets $M^\lambda \subset M$ for arbitrarily large $\lambda$, so the associated extender information is broader and harder to organise fine structurally. This is a major obstruction to extending the classical core model program. [/remark] The obstruction described in the remark points to a complementary strategy: instead of building one canonical inner model first, identify regions of truth that remain fixed across forcing extensions. This leads to the precise language of generic absoluteness. [definition: Generic Absoluteness] A generic absoluteness theorem states that, for a specified structure $A$ and a specified class of forcing notions, the truth of sentences in a specified language about $A$ is the same in $V$ and in the relevant forcing extensions of $V$. [/definition] Generic absoluteness is the forcing-side counterpart to inner-model canonicity. Inner model theory searches for canonical universes; generic absoluteness searches for domains of truth that forcing cannot disturb. [example: Mapping A Theorem To A Program] Suppose a theorem has the form: if there is a proper class of Woodin cardinals, then the first-order theory of $L(\mathbb R)$ is generically absolute for a specified class of forcing notions. The hypothesis means that for every ordinal $\alpha$ there is a Woodin cardinal $\delta>\alpha$, so the proof is not using one isolated large cardinal; it is using Woodin strength unboundedly high in $V$. The theorem belongs to the Woodin-cardinal program because the large-cardinal assumption is exactly the source of the embedding and extender strength used to control sets of reals. It belongs to descriptive set theory because $L(\mathbb R)$ is built from the ordinals and the [real numbers](/page/Real%20Numbers), so its theory is a theory about definable sets of reals and structures coded by reals. It belongs to inner model theory because $L(\mathbb R)$ is an inner model-like definable universe whose behavior is being compared with its behavior in forcing extensions. The generic-absoluteness conclusion can be read concretely as follows: if $\mathbb P$ is one of the allowed forcing notions and $G\subseteq\mathbb P$ is $V$-generic, then for each sentence $\varphi$ in the specified language, \begin{align*} L(\mathbb R)^V\models\varphi \quad\text{if and only if}\quad L(\mathbb R)^{V[G]}\models\varphi . \end{align*} Thus forcing may add new sets and new reals to the ambient universe, but the theorem says that the chosen theory of $L(\mathbb R)$ does not change across the allowed extensions. The example shows how one statement can simultaneously express large-cardinal strength, descriptive-set-theoretic control of reals, and inner-model-style stability. [/example] This example shows how a single theorem can belong to several modern programs at once. The resulting lesson is useful enough to record explicitly: HOD analysis, supercompact inner models, and generic absoluteness are not separate endpoints, but interacting tests of canonicity and stability. [explanation: Modern Inner Model Program Summary] The modern canonical inner model program seeks fine-structural inner models compatible with strong large cardinals, comparison principles for those models, and applications to HOD analysis, determinacy, and generic absoluteness. This summary is not a theorem in the formal sense; it records the organising target of a research program. The course has supplied the vocabulary needed to parse the hypotheses and conclusions of papers in this area. [/explanation] ## Summary of the Hierarchy and Further Study How should the large-cardinal hierarchy be remembered after the technical details fade? The most useful summary is not a list of names, but a list of powers: closure, reflection, compactness, embeddings, determinacy strength, and absoluteness. Each new cardinal strengthens one or more of these powers. [explanation: The Hierarchy As A Ladder Of Methods] Inaccessible cardinals provide closed initial segments $V_\kappa$ satisfying ZFC. Mahlo and indescribable cardinals strengthen reflection by ensuring that largeness reappears stationarily often below the cardinal. Weak compactness connects reflection to trees and partition properties. Measurable cardinals introduce nonprincipal $\kappa$-complete ultrafilters and ultrapower embeddings. Strong, Woodin, and supercompact cardinals extend the embedding viewpoint until it controls large fragments of the universe and substantial parts of the theory of forcing. The hierarchy is not merely linear in practical use. Some arguments need compactness, some need extender comparison, some need determinacy strength, and some need closure under long sequences. Knowing which feature a theorem uses is often more important than remembering the exact position of every named cardinal. [/explanation] A compact way to read research statements is to separate the large-cardinal hypothesis, the mathematical arena, and the stability conclusion. [example: Reading A Research Theorem] Consider the statement: assuming a proper class of Woodin cardinals, every universally Baire set of reals has forcing-invariant membership relations in the relevant extensions. The hypothesis means that for every ordinal $\alpha$ there is a Woodin cardinal $\delta>\alpha$, so the theorem is using Woodin strength unboundedly high in the universe, not just one large cardinal. The object being tested is a set of reals $A\subseteq\mathbb R$. Saying that $A$ is universally Baire means, informally, that $A$ has representations by trees whose projections continue to represent $A$ correctly after forcing. Thus if $\mathbb P$ is one of the forcing notions under consideration and $G\subseteq\mathbb P$ is $V$-generic, the intended stability assertion is that the interpretation of membership in $A$ agrees between $V$ and $V[G]$ for the reals under discussion: \begin{align*} x\in A^V \quad\text{if and only if}\quad x\in A^{V[G]} . \end{align*} Here the superscripts emphasize that the same defining representation is being read in two universes. The conclusion is therefore a generic-absoluteness statement about sets of reals: forcing may enlarge the ambient universe, but for universally Baire sets the relevant membership facts remain fixed across the allowed extensions. [/example] The prerequisites for further study divide into four strands. Inner model theory requires fine structure, mice, extenders, iteration trees, and comparison lemmas. Descriptive set theory requires projective sets, scales, determinacy, and regularity properties. Forcing axioms and generic absoluteness require iterated forcing, proper forcing, stationary-set-preserving forcing, and Boolean-valued models. Large-cardinal theory itself requires fluency with ultrapowers, normal measures, extenders, and elementary embeddings. [remark: How To Continue] Jech's text is the broad reference for ZFC, forcing, and standard combinatorics. Kanamori's book is the main guide to the higher large-cardinal hierarchy and its history. The Handbook of Set Theory is the research-level reference for inner model theory, forcing axioms, and large cardinals. Kechris provides the descriptive set theory background needed for determinacy and sets of reals. [/remark] ## Beyond and Connected Topics These notes sit between several other Androma paths. [Set Theory I](/page/Set%20Theory%20I) supplies the baseline language of ordinals, cardinals, ranks, and transfinite recursion; those tools are used here whenever the hierarchy is measured by $V_\kappa$, cofinality, or closure under power sets. [Set Theory II: Forcing](/page/Set%20Theory%20II%3A%20Forcing) is the natural continuation on the forcing side: large cardinals become especially useful when one asks which embeddings can be lifted through forcing extensions and which truths remain absolute afterward. For readers wanting a broader foundations route, [Cambridge II Logic and Set Theory](/page/Cambridge%20II%20Logic%20and%20Set%20Theory) gives a compact bridge from first-order logic and axiomatic set theory to the model-theoretic viewpoint used in reflection and absoluteness arguments. [Model Theory I: Foundations](/page/Model%20Theory%20I%3A%20Foundations) and [Model Theory II: Types and Stability](/page/Model%20Theory%20II%3A%20Types%20and%20Stability) develop the surrounding language of elementary equivalence, structures, and definability, which helps explain why large-cardinal embeddings are so powerful despite not being ordinary set-sized homomorphisms. The determinacy chapters connect large cardinals to analysis through regularity properties of sets of reals. For that direction, [Measure Space](/page/Measure%20Space) and [Lebesgue Measure](/page/Lebesgue%20Measure) provide the measure-theoretic background behind measurability conclusions, while [Axiom of Choice](/page/Axiom%20of%20Choice) explains the choice principle whose full strength is incompatible with determinacy. Together these connected topics show why large cardinal theory is not isolated from the rest of mathematics: it organizes model existence, forcing stability, definability, and regularity into one consistency-strength hierarchy. The conceptual endpoint of the course is the following: large cardinals are not isolated assertions that certain cardinals exist. They are principles saying that the universe has enough height, closure, and self-similarity to support powerful forms of reflection and stability. The modern search for canonical universes asks whether those principles can be organised into a coherent picture of the set-theoretic universe itself. ## References - Thomas Jech, *Set Theory*, especially the chapters on large cardinals, forcing, and stationary sets. - Akihiro Kanamori, *The Higher Infinite*, for the large-cardinal hierarchy, indescribability, measurability, Woodin cardinals, and historical context. - Kenneth Kunen, *Set Theory*, for forcing, elementary embeddings, ultrapowers, and foundational constraints around large-cardinal arguments. - Matthew Foreman and Akihiro Kanamori, editors, *Handbook of Set Theory*, for research-level treatments of inner model theory, determinacy, forcing axioms, and large cardinals. - Alexander S. Kechris, *Classical Descriptive Set Theory*, for projective sets, regularity properties, and the descriptive-set-theoretic background behind determinacy.