Towards beating the probabilistic lower bound on $q$-perfect hashing for all $q$

For an integer $q\ge 2$, a perfect $q$-hash code $C$ is a block code over $[q]:=\{1,\ldots,q\}$ of length $n$ in which every subset $\{\mathbf{c}_1,\mathbf{c}_2,\dots,\mathbf{c}_q\}$ of $q$ elements is separated, i.e., there exists $i\in[n]$ such that $\{\mathrm{proj}_i(\mathbf{c}_1),\dots,\mathrm{proj}_i(\mathbf{c}_q)\}=[q]$, where $\mathrm{proj}_i(\mathbf{c}_j)$ denotes the $i$th position of $\mathbf{c}_j$. Finding the maximum size $M(n,q)$ of perfect $q$-hash codes of length $n$, for given $q$ and $n$, is a fundamental problem in combinatorics, information theory, and computer science. In this paper, we are interested in asymptotical behavior of this problem. More precisely speaking, we will focus on the quantity $R_q:=\limsup_{n\rightarrow\infty}\frac{\log_2 M(n,q)}n$. A well-known probabilistic argument shows an existence lower bound on $R_q$, namely $R_q\ge\frac1{q-1}\log_2\left(\frac1{1-q!/q^q}\right)$ \cite{FK,K86}. This is still the best-known lower bound till now except for the case $q=3$ for which K\"{o}rner and Matron \cite{KM} found that the concatenation technique could lead to perfect $3$-hash codes beating this the probabilistic lower bound. The improvement on the lower bound on $R_3$ was discovered in 1988 and there has been no any progress on lower bound on $R_q$ for more than $30$ years despite of some work on upper bounds on $R_q$. In this paper we show that this probabilistic lower bound can be improved for $q$ from $4$ to $15$ and all odd integers between $17$ and $25$, and \emph{all sufficiently large} $q$. Although we are not able to prove that our construction can beat the probabilistic method for all $q$, the fact that our construction beat the probabilistic method for both small and large $q$ sheds light on that this construction might hold for all $q$.