Layered King Summation on Sequences and Matrices: Exact Behavior of the Standard Deviation
We study a layered summation process, which we call the king model, on sequences and matrices. On a sequence there are three operations: type 1 adds zero, type 2 adds every entry to itself, and type 3 adds one entry (the "king") to all entries, every entry becoming king in turn. On an m × n matrix the kings are the rows. Using only the translation invariance of the standard deviation and the splitting of the variance into a within-column and a between-column part, we prove the following. (i) On a sequence of length n, type 3 preserves the standard deviation exactly, the mean becomes n + 1 times larger after one round, and the sequence can be recovered from its last layer. (ii) On a matrix, one round is the map A → (I + J)A with J the all-ones matrix; the within-column variance is unchanged while the between-column variance is multiplied by (m + 1)², so the standard deviation at the end of a round equals the initial one if and only if all column sums are equal. (iii) The standard deviation is the same on every layer if and only if every row is constant, which refutes the natural guess that only matrices with all entries equal have this property. (iv) The process may be continued indefinitely: for sequences the standard deviation never changes, whichever rule is used to continue the kings, while for matrices the standard deviation after k rounds is again determined by the within-column variance and the column sums, and grows geometrically (kings taken from the latest layer) or linearly (original kings repeated) unless all column sums are equal. The results rest on classical identities and we claim no new tools; the contribution is the formulation, exact formulas and "if and only if" characterizations. We illustrate their use as test cases for the numerical stability of variance formulas and as an exactly solvable toy model of variance growth under a residual connection followed by sum pooling. Every statement is checked numerically, including exhaustive searches over integer matrices.
Authors
- Jaka Perdana
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24269435
- Primary Topic
- Matrix Theory and Algorithms
- Type
- preprint