The study of number fields involves objects called class groups, which measure how far a number system strays from having unique factorization. A long-standing set of predictions called the Cohen-Lenstra-Gerth conjectures describes how class groups should be distributed randomly across large families of number fields. This paper focuses on imaginary quadratic fields, which are number systems built by adjoining the square root of a negative number to the ordinary rationals. For each squarefree integer d, you get two naturally related fields by taking square roots of negative d and of negative d times a fixed number d0. These two fields share a lot of arithmetic structure, so a key question is whether their class groups behave independently despite being correlated.
The main result is that even though these two fields are arithmetically intertwined, a specific part of their class groups, called the 4-rank, becomes statistically independent as d grows large. More precisely, the joint distribution of the 4-ranks of the two fields converges to what you would expect if the two class groups were chosen completely independently, each according to the Cohen-Lenstra-Gerth distribution. The paper also goes further by studying a larger number field built from both square roots simultaneously, called a biquadratic field. Under a mild extra condition, the class group of this bigger field decomposes cleanly as a direct sum of pieces coming from the two smaller fields, which lets the authors predict the full distribution of the class group of the biquadratic field.
The proofs rely on combining two technical tools. The first is a method developed by Alexander Smith called the box method, which organizes information about how prime ideals behave across the family of fields. The second is a new quantitative technique for inverting certain algebraic moment conditions involving structured matrices called Redei matrices, which encode the arithmetic of 2-power divisibility in class groups. Together these tools let the authors extract precise statistical independence from what appears to be a strongly correlated setup, and they make the error in the approximation explicitly quantifiable in terms of how large d is allowed to grow.