Naming and terminology¶
Hifuku draws on two fields: phylogenetics, and quality-diversity search. Because quality-diversity grew out of evolutionary computation, it borrows biology's own words, so the danger here is the reverse of the usual cross-field problem: the two fields often use the same word for different ideas. This page fixes the vocabulary. Write for a biologist first, keep the biological meaning as the default, and mark the algorithmic meaning where a word could be read both ways.
Two rules hold in code and documentation:
- Use one name for one thing.
- Do not call a search-steering quantity a likelihood or a fitness landscape. The elevation of the map is always the measured alignment log-likelihood of a tree, computed directly. The survey machinery only chooses which trees to evaluate.
Three quantities¶
Hifuku computes a likelihood landscape for a gene and renders it over a shared 2D chart. Three quantities do three separate jobs.
| Quantity | Definition | Role |
|---|---|---|
| log-likelihood | log P(alignment given the tree and model) | rendered as the elevation of the map |
| the elite archive | the best tree per niche, filled by the select-vary-place survey | selects which trees to evaluate |
| the map | log-likelihood over the 2D chart | the rendered result |
The log-likelihood is the scientific result. The survey selects samples. The map takes its elevation from the log-likelihood, not from the search.
The same words, two meanings¶
Quality-diversity search names its parts after biology. In this project the biological meaning is the default; the algorithmic meaning is used only for the mechanics of the search.
| Word | The science (default) | The search (MAP-Elites) |
|---|---|---|
| evolution | the substitution process along a tree | the mutate-and-select loop over candidate trees |
| mutation | a modeled sequence substitution | a tree move: NNI, SPR, branch slide, branch scale |
| fitness, performance | reproductive success (not modeled here) | the alignment log-likelihood of a candidate tree |
| population | the taxa and sequences under study | the archive of elite trees |
| parent, offspring | ancestor and descendant organisms | a candidate tree and the varied copy made from it |
| lineage | a line of descent among organisms | the chain of successive elite variants in the search |
| niche | an ecological niche | a cell of the feature space (the 2D chart) |
One rule carries the convention: if a quantity is the measured log-likelihood or part of the substitution model, it is the science; if it only steers the search, it is the algorithm.
Phylogenetics vocabulary¶
These terms fix Hifuku's usage within phylogenetics, and mark a few words a reader from enhanced sampling or MCMC might otherwise reach for.
| Use this | Not this | Note |
|---|---|---|
| log-likelihood, elevation | free energy | Hifuku measures the log-likelihood directly; it does not reconstruct a free energy |
| chart coordinate | collective variable | the barycentric coordinate on the chart, not an enhanced-sampling CV |
| walker | posterior chain | a parallel evaluation lane, a chain of successive candidate trees, not an MCMC chain that targets a posterior |
| split | clade | trees are unrooted, so a branch is a split |
| metric saturation | substitution saturation | loss of tree-distance resolution, distinct from sequence saturation |
Notes for the reader from another field¶
From quality-diversity and evolutionary computation¶
The search borrows biology's words, but here they name the mechanics of a search over candidate trees, not claims about biological evolution. A "mutation" is a tree move, "fitness" is the alignment log-likelihood, the "population" is the archive, and a "lineage" is a chain of successive variants. The performance is a directly computed log-likelihood, not an organism's reproductive success. Hifuku is an illumination algorithm (MAP-Elites): it keeps the best tree in every niche and maps the whole landscape, rather than driving toward one optimum.
From phylogenetic MCMC¶
Hifuku does not estimate a posterior distribution. The survey is not an MCMC chain: placement is a keep-if-better rule with no acceptance ratio and no detailed balance, and it uses goal-directed moves such as long jumps and restarts. Such moves would be incorrect in a posterior sampler such as MrBayes or BEAST. They are correct here because every elevation is an exact log-likelihood, computed for each tree the survey evaluates, not a sample from a target density.
Ambiguous terms¶
- kernel: state which one you mean. A CUDA kernel is a device function; a KDE kernel is a density estimate used in offline analysis. Do not call the Shepard interpolation weight a kernel.
- chart, not space: the anchor plane is a local Euclidean chart of non-Euclidean tree space (CAT(0), or BHV), not the space itself.
- log-likelihood gate: the relative margin or absolute floor that bounds the surveyed region is a gate on which trees create niches, not a bound on tree space. It keeps the filled footprint within the reasonably-likely region.
- niche: a cell of the barycentric chart, indexed by integer coordinates. The grid is unbounded, so every chart point has a niche; the filled niches form the map's footprint.