Even if macros and for loops "contaminate" the declarativeness and are unavoidable at scale, it's very easy to analyze the core of spreadsheet logic from the referentially transparent point of view of dataflow programming. The problem is that the direct view of a spreadsheet does not show this layer, even if it is what people program.

Spreadsheets are a small, mostly applicative language with references. We can ignore the "grid" presentation as it's just an easy way to examine the values of each of these references. Instead, we just say that a spreadsheet program is a collection of named formulae in this language where the references each refer to the names of one of the formulae. A program is well-formed if all references have a target.

Evaluation of a spreadsheet language involves seeking a fixed-point of this collection of formulae. If the formulae are not well-formed then their evaluation may become "stuck". If all references are targeted then the spreadsheet may diverge. Generally, both of those cases are uninteresting and we seek convergent programs.

Spreadsheet programs are finally displayed by iterating their formulae to convergence and displaying the result of every formula at once in a grid.

---

If you think about spreadsheets as having the denotation I just described then they are referentially transparent, have a well-defined notion of "variable" as opposed to "assignable", have (restricted) beta equivalence, have a denotational semantics, talk about "what, not how", and, if you do some analysis to find the connected components of your dataflow graph, are implicitly parallelizable.