Let us write $\{ A_1, \ldots, A_k\} \rightrightarrows^n \{B_1, \ldots, B_m\}$ for a relation $R$ if for all $\langle a_1, \ldots, a_k\rangle$ we have $\left\lvert \pi_{B_1, \ldots, B_m} \sigma_{\bigwedge_{i = 1}^k A_i = a_i} R\right\rvert \le n$ (and $A_i$, $B_j$ are attributes from the schema of $R$), i.e, for any choice of values for the attributes $A_i$ at most $n$ corresponding tuples of values for $B_j$ can be found in $R$. Then $\{A_1, \ldots, A_k\} \rightrightarrows^1 \{B_1, \ldots, B_m\}$ is just the functional dependency $\{A_1, \ldots, A_k\} \rightarrow \{B_1, \ldots, B_m\}$, and $\emptyset \rightrightarrows^n \{B_1, \ldots, B_m\}$ is an upper bound on the number of possible $B_j$-tuples occurring in $R$ (regardless of any selection).
One may come up with a bunch of possible inference rules similar to Armstrong's axioms, for example
- reflexivity: $\mathcal{A}' \subseteq \mathcal{A} \implies \mathcal{A} \rightrightarrows^1 \mathcal{A}'$,
- augmentation: $\mathcal{A} \rightrightarrows^{n} \mathcal{B} \implies \mathcal{A} \cup \mathcal{C} \rightrightarrows^{n} \mathcal{B} \cup \mathcal{C}$,
- transitivity: $\mathcal{A} \rightrightarrows^{n_1} \mathcal{B}$, $\mathcal{B} \rightrightarrows^{n_2} \mathcal{C} \implies \mathcal{A} \rightrightarrows^{n_1 n_2} \mathcal{C}$,
- decomposition: $\mathcal{A} \rightrightarrows^{n} \mathcal{B} \cup \mathcal{C} \implies \mathcal{A} \rightrightarrows^{n} \mathcal{B}$, $\mathcal{A} \rightrightarrows^{n} \mathcal{C}$,
- composition: $\mathcal{A}_1 \rightrightarrows^{n_1} \mathcal{B}_1$, $\mathcal{A}_2 \rightrightarrows^{n_2} \mathcal{B}_2 \implies \mathcal{A}_1 \cup \mathcal{A}_2 \rightrightarrows^{n_1 n_2} \mathcal{B}_1 \cup \mathcal{B}_2$.
Indeed, by setting $n = n_1 = n_2 = 1$, we obtain Armstrong's axioms.
Let $\Sigma = \{\mathcal{A}_i \rightrightarrows^{n_i} \mathcal{B}_i\}_{i = 1}^N$ be a set of such "cardinality constraints". Again following the conventions from functional dependencies, let us write $\Sigma \vDash \mathcal{A} \rightrightarrows^n \mathcal{B}$ if $\mathcal{A} \rightrightarrows^n \mathcal{B}$ holds in every relation $R$ where all constraints in $\Sigma$ hold. Let $\min_\Sigma(\mathcal{A}, \mathcal{B})$ be the smallest $n$ such that we have $\Sigma \vDash \mathcal{A} \rightrightarrows^n \mathcal{B}$. How can we compute (or bound) $\min_\Sigma(\mathcal{A}, \mathcal{B})$? Is there a way to minimize $\Sigma$ and obtain (the smallest) $\Gamma$ such that $\min_\Gamma(\mathcal{A}, \mathcal{B}) = \min_\Sigma(\mathcal{A}, \mathcal{B})$ for all sets of attributes $\mathcal{A}$, $\mathcal{B}$?
I'd be interested in references or algorithms regarding the constraints defined above. Have they been studied? What is the standard terminology or notation?
