Large Cardinals & Inner Models

Set theory does not stop at the infinite cardinals $\aleph_0, \aleph_1, \aleph_2, \ldots$ that arise from the ZFC axioms alone. By postulating the existence of cardinals so enormous that their mere existence cannot be proved in ZFC, mathematicians have uncovered a remarkably coherent hierarchy that calibrates the strength of mathematical theories and illuminates the fine structure of the set-theoretic universe. These large cardinal axioms and the inner models built to accommodate them form the backbone of modern set theory, connecting foundational questions about consistency, determinacy, and definability into a single sweeping narrative.

Inaccessible and Mahlo Cardinals

The story of large cardinals begins with a deceptively simple question: can a cardinal be so large that the sets below it already form a model of all of set theory? A cardinal $\kappa$ is strongly inaccessible if it satisfies three conditions: it is uncountable, it is regular (meaning no collection of fewer than $\kappa$ sets, each of size less than $\kappa$, can have a union of size $\kappa$), and it is a strong limit (meaning $2^\lambda < \kappa$ for every $\lambda < \kappa$). The weaker notion of weakly inaccessible, where the strong limit condition is replaced by the requirement that $\kappa$ is a limit cardinal, was introduced by Felix Hausdorff in his pioneering 1908 work on order types and cardinal arithmetic.

The key consequence of strong inaccessibility is that the cumulative hierarchy truncated at $\kappa$, written $V_\kappa$, is itself a model of all the ZFC axioms. The replacement axiom holds because regularity prevents any function defined on a set of size less than $\kappa$ from reaching beyond $V_\kappa$, and the power set axiom holds because the strong limit condition ensures that power sets of sets in $V_\kappa$ remain in $V_\kappa$. By Godel’s second incompleteness theorem, this means ZFC cannot prove that inaccessible cardinals exist — for if it could, it would prove its own consistency, contradicting the incompleteness theorem (assuming ZFC is consistent). Thus already at the first rung of the large cardinal ladder, we encounter axioms that genuinely transcend the base theory.

Mahlo cardinals, introduced by Paul Mahlo in 1911, impose a stronger reflection condition. A cardinal $\kappa$ is Mahlo if it is inaccessible and the set of inaccessible cardinals below $\kappa$ is stationary in $\kappa$ — that is, it meets every closed unbounded subset of $\kappa$. Since stationary sets are in a precise sense “large,” a Mahlo cardinal sits atop a rich population of inaccessible cardinals, each of which itself provides a model of ZFC. One can iterate the idea: a hyper-Mahlo cardinal requires that the set of Mahlo cardinals below it be stationary, and so on through a transfinite hierarchy of reflection principles. Each level in this hierarchy strictly exceeds the consistency strength of all the levels below it, yet all of them remain comparatively modest in the full landscape of large cardinals.

Measurable Cardinals and Ultrafilters

A dramatic leap in consistency strength occurs with measurable cardinals, which bring measure-theoretic and model-theoretic ideas into the combinatorial world of large cardinals. A cardinal $\kappa$ is measurable if there exists a $\kappa$-complete non-principal ultrafilter on $\kappa$ — that is, a $\{0,1\}$-valued measure defined on all subsets of $\kappa$ that assigns measure zero to every singleton and is closed under intersections of fewer than $\kappa$ sets. The concept was introduced by Stanislaw Ulam in 1930 in his investigation of the measure problem, the question of whether a countably additive two-valued measure can exist on any set.

The importance of measurable cardinals exploded in 1961 when Dana Scott proved a landmark result: if a measurable cardinal exists, then $V \neq L$. This was the first concrete evidence that Godel’s constructible universe $L$ is “too thin” to capture the full combinatorial richness of the set-theoretic universe. Scott’s proof proceeds through the ultrapower construction. Given a $\kappa$-complete non-principal ultrafilter $U$ on $\kappa$, one builds the ultrapower $\text{Ult}(V, U)$, and Los’s theorem guarantees that the canonical embedding

$j : V \to M$

is an elementary embedding — it preserves all first-order properties. The critical point of $j$, the least ordinal moved, is exactly $\kappa$, and $j(\kappa) > \kappa$. The transitive collapse $M$ of the ultrapower is an inner model that agrees with $V$ on sets of rank less than $\kappa$ but diverges above. In $L$, no such embedding can exist because the rigidity of the constructible hierarchy prevents the existence of the required ultrafilter.

The fine structure of measurable cardinals is further organized by the Mitchell order, which partially orders the normal measures on $\kappa$ by the relation $U \triangleleft W$ if and only if $U$ belongs to the ultrapower of $V$ by $W$. The Mitchell rank of a measurable cardinal — how tall this partial order can be — provides a refined measure of its strength, ranging from $1$ (a single normal measure) up through the ordinals, and plays a central role in inner model theory.

Woodin Cardinals and Determinacy

Between measurable and supercompact cardinals lies a class of large cardinals whose significance is not purely combinatorial but stems from deep connections to the descriptive set theory of the real line. A cardinal $\delta$ is a Woodin cardinal if for every function $f : \delta \to \delta$ there exists a cardinal $\kappa < \delta$ that is closed under $f$ and carries an elementary embedding $j : V \to M$ with critical point $\kappa$ such that $V_{j(f)(\kappa)} \subseteq M$. The definition is technical, but the payoff is profound.

The central achievement connecting large cardinals to the structure of definable sets of reals is the Martin-Steel theorem (1989): if there exist infinitely many Woodin cardinals with a measurable cardinal above them all, then Projective Determinacy (PD) holds. Projective Determinacy asserts that every projective set of reals — every set obtainable from open sets by complementation and continuous images, iterated finitely many times — is determined, meaning that in the associated infinite two-player game, one of the players has a winning strategy. Donald Martin had proved Borel determinacy in ZFC alone in 1975, but extending determinacy to the projective sets required the heavy artillery of large cardinals.

Why does determinacy matter? Because it yields a complete and beautiful structure theory for the projective sets. Under PD, every projective set of reals is Lebesgue measurable, has the property of Baire, and satisfies the perfect set property (so the continuum hypothesis holds for projective sets). These are exactly the regularity properties that pathological sets constructed via the axiom of choice can violate. W. Hugh Woodin further deepened the connection by showing that large cardinal axioms not only imply determinacy but are in a precise sense equivalent in consistency strength to determinacy principles, establishing a bridge between the combinatorics of enormous cardinals and the analysis of definable subsets of $\mathbb{R}$. This work, developed throughout the 1980s and 1990s by Woodin, Martin, John Steel, and others, transformed the field by revealing that questions about real numbers and questions about the heights of the set-theoretic universe are intimately intertwined.

The Large Cardinal Hierarchy

One of the most striking empirical discoveries in modern set theory is that large cardinal axioms are linearly ordered by consistency strength. If axiom $A$ implies $\text{Con}(\text{ZFC} + B)$ — that is, the consistency of $B$ — then $A$ is said to have greater consistency strength than $B$. Despite the enormous variety of large cardinal definitions, every pair that has been studied stands in this linear order. The hierarchy, in ascending order of consistency strength, includes:

$\text{inaccessible} < \text{Mahlo} < \text{weakly compact} < \text{Ramsey} < \text{measurable} < \text{strong} < \text{Woodin} < \text{supercompact} < \text{extendible} < \text{huge}$

Beyond huge cardinals lie the rank-into-rank axioms, labeled $I3$, $I2$, $I1$, and $I0$ in decreasing order of restrictiveness. The axiom $I3$ posits an elementary embedding $j : V_\lambda \to V_\lambda$ for some limit ordinal $\lambda$; $I1$ posits $j : V_{\lambda+1} \to V_{\lambda+1}$; and $I0$ posits such an embedding that is moreover iterable. These axioms push consistency strength to its very edge.

At the top of the hierarchy sits the Reinhardt cardinal, defined as the critical point of an elementary embedding $j : V \to V$. Kenneth Kunen proved in 1971 that Reinhardt cardinals are inconsistent with the axiom of choice: no such embedding can exist if AC holds. This result, known as Kunen’s inconsistency, marks a definitive boundary to the large cardinal program (at least in the presence of choice). Just below this boundary, Vopenka’s principle asserts that for any proper class of structures of the same type, one elementally embeds into another; it is equivalent in strength to the existence of various large cardinals and serves as a powerful organizing axiom in category theory and abstract model theory.

A supercompact cardinal $\kappa$ is one for which, for every $\lambda \geq \kappa$, there exists an elementary embedding $j : V \to M$ with critical point $\kappa$ such that $j(\kappa) > \lambda$ and $M$ is closed under $\lambda$-sequences (that is, ${}^\lambda M \subseteq M$). Supercompacts are strong enough to imply a wealth of combinatorial consequences and serve as the engine behind many independence results, yet their consistency strength falls well below the rank-into-rank axioms. The linearity of this entire hierarchy remains an unexplained phenomenon — no one has proved that it must be linear, yet no counterexample has ever been found. It suggests a deep structural coherence in the space of mathematical theories that we do not yet fully understand.

Inner Models and Core Model Theory

While large cardinal axioms expand the universe of sets upward, inner model theory works in the opposite direction: it constructs the thinnest possible universes that can still accommodate specific large cardinals, thereby calibrating their exact consistency strength. The prototype is Godel’s constructible universe $L$, built by iterating definable power set operations through the ordinals. At each successor stage $\alpha + 1$, one forms $L_{\alpha+1}$ by collecting only those subsets of $L_\alpha$ that are first-order definable over $L_\alpha$ with parameters, rather than taking the full power set. The result is a proper class $L = \bigcup_\alpha L_\alpha$ in which every set is ordinal-definable — constructible from the ordinals by a fixed list of operations that Godel identified.

The constructible universe satisfies both the axiom of choice and the generalized continuum hypothesis ($2^{\aleph_\alpha} = \aleph_{\alpha+1}$ for all $\alpha$), and it is the smallest inner model of ZFC that contains all the ordinals. The condensation lemma — which states that every elementary submodel of an $L_\alpha$ is isomorphic to some $L_\beta$ — is the key structural property of $L$ and the engine behind most of its good behavior. However, as Scott’s theorem shows, $L$ cannot contain a measurable cardinal, so it is inadequate as a model for stronger axioms.

To build inner models for measurable cardinals, one passes to $L[U]$, the constructible universe relativized to a normal measure $U$ on $\kappa$. The model $L[U]$, studied extensively by Kunen and Silver in the 1970s, satisfies ZFC together with “there is exactly one measurable cardinal,” and it shares many of the fine-structural properties of $L$, including a form of condensation and GCH. The core model $K$, developed by Ronald Jensen and later by Anthony Dodd, Jensen, and John Steel, goes further: it is the largest inner model that does not contain the large cardinal in question, and it approximates $V$ as closely as possible without surpassing a given consistency strength threshold.

The bridge between inner models and the surrounding universe is the covering lemma. In its original form, due to Jensen, it states: if the sharp $0^\#$ does not exist (equivalently, if there is no non-trivial elementary embedding $j : L \to L$), then for every uncountable set $X$ of ordinals there is a constructible set $Y \supseteq X$ with $|Y| = |X|$. In other words, $L$ “covers” $V$ in the sense that it can approximate arbitrary sets of ordinals without increasing their cardinality. When $0^\#$ does exist, the covering lemma fails and $L$ is far from $V$, but one can then pass to a richer core model and recover a covering lemma at a higher level. This interplay — building inner models, proving covering lemmas, and identifying the exact point where they break down — has been extended by Steel, Mitchell, and others up through Woodin cardinals, though constructing canonical inner models at the level of supercompact cardinals remains one of the outstanding open problems of contemporary set theory, as explored in The Continuum Hypothesis and throughout the broader study of forcing and independence.