[GordonYuanyc]

Design doc: docs/sketch_analytics.md defines the theoretical error, memory and time analytics asap_sketchlib provides for each sketch, as big O and as callable formulas, and the SketchAnalytics trait (theoretical_memory_bytes, theoretical_update_ops, theoretical_query_ops, theoretical_merge_ops, theoretical_error, params_for) that exposes them. Each sketch section maps its symbols to the paper's and to the code's parameter names. Empirical measurements belong to sketch-bench.

Only the doc changes; no code changes.

Where ASAPPlanner's formulas differ today

• Kmv update is 1 op and memory is 8k bytes; the update scans the kept hashes linearly, and the heap holds up to about 2k slots (KMV update cost is not one operation per row ASAPPlanner#945).
• Count-Min and Count Sketch with heap update is d + 1; it also re-estimates the key and sifts the heap.
• A heap slot is 24 bytes; theoretical_memory_bytes counts the entry, digest, index and key bytes.
• UnivMon update is λ(L)·(d + 1); it also writes row sums, re-estimates the key and updates each layer's heap.
• Kll update is 1; compaction sorts the bottom level, lg(c_b + 1) on average.
• Kll memory leaves out merge scratch space and the CDF cache.
• DdSketch update is 1; it touches six cells.
• DdSketch memory grows stores in 128-bin chunks; theoretical_memory_bytes uses 2(s + 126) slots per store.
• A merge is 1 op per input state, whatever the sketch's size.
• HyperLogLog parameters do not name the estimator, so the Classic bounded contract is also applied to Ertl.
• HyperLogLog sizing returns precisions asap_sketchlib does not build by default, such as 15 and 17.
• The Kll contracts apache_datasketches_kll_empirical_99_a9b42755072b and ..._delta_scaled_... become kll_quantile_empirical_99_v1 and kll_quantile_empirical_99_delta_scaled_v1, since the fit is measured on asap_sketchlib.

Open questions

The two at the end of the doc: the Kll rank() contract, and CountSketch sizing under ASAPv1's 20-row limit.

🤖 Generated with Claude Code