SuperBEST v5.32026-09-13arXiv:2603.2185223 operators catalogued · no 24th found (conjecture)

SuperBEST Routing Table

Best known F16-node constructions for every elementary arithmetic primitive. Every count is an upper bound; none is shown to be minimal. Taxonomy: 23 operators. CONJ_NO_OP_24, that no 24th operator passes the PGC+AIT filter on an open domain, is a conjecture: the monogate repository's exploration notes argue it by case analysis and a filter pass over the remaining candidate operators, and no Lean proof exists.

Two-layer accounting policy — Layer 1 (F16)formal theorems, arXiv paper; which sixteen operators each table counts is said under the headline.Layer 2 (23-op)library/ML/physics use; extended operators count as 1n. Must be labeled.
15n / 79.5% savings (positive domain, 10-op headline vs 73n naive, F16)

Which F16. The core table counts nodes in the sixteen operators on monogate.org/framework ↗ and nothing else, since monogate.org recounted it on 2026-09-13. ln x takes 2 nodes there, LEdiv(0, F13(−1, x)), so the positive-domain total is 15n / 79.5%. Counting ln x = EXL(0, x) as one node instead, with the census operator EXL, exp(x)·ln y, gives 14n / 80.8%. Both totals are upper bounds. Each construction is evaluated numerically on its domain by the monogate repository's python/tests/test_superbest_f16_constructions.py, which this site does not run, and this site gates no Lean theorem that states either total. Before each deploy the totals are compared with python/monogate/superbest.py, and every node count in the table with monogate.org's copy. The high-impact tables below follow the taxonomy-lock paper, whose "F16 orbit" is a different sixteen that leaves out LEAd (F11 on /framework).

General domain. The general basket is exp, neg, add, sub, mul and div, each with one F16 tree valid for every real input where the operation is defined (div needs y ≠ 0). mul takes 3 nodes, LEdiv(0, F13(y, DEML(x, 1))), and div takes 8. A numerical search over F16 trees with constant leaves 0, 1, −1, 2 and 1/2 finds no smaller signed division (python/benchmarks/superbest_f16/search.pyin the monogate repository); a search is not a proof. Before the recount, mul and div counted sign-dispatch case splits, one circuit per sign quadrant, not trees. Two operations left the basket: abs, because no real F16 tree computes |x| (every tree is real-analytic where it is defined, and |x| is not analytic at 0; a paper argument, not a Lean proof), and ln, which has no real value for x ≤ 0. recip has a row-level entry for x ≠ 0; sqrt and pow have none over all reals. The naive costs are the monogate library's pure-EML figures, not re-derived in the recount.

Core Arithmetic Primitives (F16 Layer 1)

OperationNodes (x>0)v4 costConstructionNotes
exp(x)1n1EML(x,1) = e^xprimitive — no domain restriction
ln(x)2n1LEdiv(0, F13(-1,x)) = 0 - ln(x^-1) = ln(x), x > 0No single F16 node equals ln x (a search over one-node trees finds none). EXL(0,x) = ln(x) in one node, but EXL is a census operator outside F16; the count with EXL on this page uses it.
recip(x) = 1/x1n1F13(-1,x) = exp(-ln(x)) = 1/x, x > 0R16-C1's ELSb(0,x) is the same function at the same count for x > 0 (ELSb is a census operator). The general entry is row-level; recip is not in the general basket.
div(x,y) = x/y2n2F16fn(x, F13(-1,y)) = exp(ln(x) + ln(1/y)) = x/y, x > 0, y > 0Recounted in F16 on 2026-09-13: one tree for every x and y != 0. The earlier general entry counted a sign-dispatch case split.
neg(x) = -x2n2LEdiv(0, EML(x,1)) = 0 - ln(exp(x)) = -x; EML(x,1) = exp(x) > 0, so LEdiv is always validT09 wrote the same count with the census operator EXL, as EXL(0, DEML(x,1)). No single F16 node computes -x (search over one-node trees).
mul(x,y) = x*y1n2F16fn(x, y) = exp(ln(x) + ln(y)) = x*y for x, y > 0One tree over all reals since the F16 recount of 2026-09-13.
sub(x,y) = x - y2n2LEdiv(x, EML(y,1)) = ln(exp(x)/exp(y)) = x - yT33; 2n for all reals; X8 confirmed
add(x,y) = x + y2nLEdiv(x, DEML(y,1)) = ln(exp(x)/exp(-y)) = x + yNEW v5 (ADD-T1): all real x,y. Replaced add_pos=3n and add_gen=11n with unified 2n.
sqrt(x) = x^(1/2)1n2F13(0.5, x) = exp(0.5·ln(x)) = sqrt(x), x > 0 — single F16 node (EPL(0.5, x) in the census)T_SQRT_1N; EPL(0.5,x) = ELMl direct primitive. Same mechanism as pow=1n. Corrected from v5.1 (was 2n). Spot values: EPL(0.5,4)=2.0, EPL(0.5,9)=3.0, EPL(0.5,16)=4.0.
pow(x,n) = x^n1n3F13(n, x) = exp(n·ln(x)) = x^n, x > 0 — single F16 node (EPL/ELMl in the census)X20 RESOLUTION: EPL/ELMl = exp(x*ln(y)) = y^x IS in F16 census. pow(x,n)=x^n=EPL(n,x) for x>0 costs 1n (direct), not 3n. The 3n construction used EXL+ELAd+EML explicitly rather than recognizing EPL as a primitive. For positive domain: corrected to 1n.

Extended Operators (23-op Beyond F16)

Cat A: genuine F16 algebraic shortcut. Cat B: genuine via EML sign-variant. Cat C: notation-only; with no F16 shortcut, the saving exists only in the 23-op count. Node counts here and in the high-impact tables are from the SuperBEST taxonomy lock of 2026-04-21. Nothing on this site re-derives them; before each deploy they are compared cell by cell with the copy monogate.org shows.

OpF16 cost (L1)23-op cost (L2)Genuine savingCategoryDescription
EEM3n1n1AAlgebraic shortcut (add+exp)
EED3n1n1AAlgebraic shortcut (sub+exp)
EES1n/4n1n3n(const)/0n(gen)B/CGenuine F16 for const arg; notation-only general
LLA3n1n1AAlgebraic shortcut (mul+ln)
LLS3n1n1AAlgebraic shortcut (div+ln)
LLD4n1n0CNotation-only — no F16 shortcut
EEA4n1n0CNotation-only — enables LSE=2n in 23-op

High-Impact Expressions

Layer 1 = F16 ground-truth; Layer 2 = 23-op extended (must be labeled)

ML / Smooth Functions

ExpressionNaiveF16 (Layer 1)23-op (Layer 2)Saving type
softplus ln(1+exp(x))4n2n2nB: EML(x,1/e)+ln
LSE(x,y) = ln(exp(x)+exp(y))8n4n2nF16: exp+exp+add+ln (corrected from 5n)
sigmoid 1/(1+exp(-x))6n4n3nF16 route
KL term l*ln(l/m)6n5n3nA: div+ln+mul

Quantum / Information

ExpressionNaiveF16 (Layer 1)23-op (Layer 2)Saving type
quantum rel entropy (per term)6n5n3nA: genuine 1n save
von Neumann entropy (per term)8n7n4nA: 1n save
Bures distance5n4n3nB: genuine
partition fn pair exp(-bEj)/exp(-bEi)10n10n7nC: EEA notation-only
spectral decay exp(-g*t)4n3n2nA: mul+exp
Kraus factor sqrt(1-exp(-g*t))6n4n3nB: EML-variant+EPL

Physics

ExpressionNaiveF16 (Layer 1)23-op (Layer 2)Saving type
Boltzmann ratio exp(-b*(Ej-Ei))10n6n5nA: sub+mul+neg+exp
Mayer f-function exp(-b*u)-16n3n3nB: EML_neg+mul
hydrogen radial decay4n3n3nB: genuine
Fermi-Dirac 1/(exp((e-m)/kT)+1)8n6n4nA: algebraic route

Special Functions (all require infinite F16 depth)

FunctionF16 / 23-op costReason
erf(x)T01: infinite zeros
J_0(x) BesselT01: infinite zeros
Gamma(x)AIL theorem
sin(x), cos(x)T01/AIL

Honest Savings Summary

CatalogLayer 1 — F16 genuineLayer 2 — 23-op extended
Core arithmetic (10 ops, positive domain)79.5%80.8%
ML functions (5 items)35%62%
Quantum (8 items)19%44%
Physics (5 items)29%40%
Special functions0% (all inf)0% (all inf)
Aggregate~12%~40%

The core-arithmetic row's Layer 1 cell is the positive-domain F16 headline; its Layer 2 cell counts ln x = EXL(0, x) as one node, as monogate.org's does. Both are computed from totals that are checked against the monogate.superbest library before each deploy. The ML, quantum and physics rows are pooled over the SuperBEST taxonomy lock's own item rows, as monogate.org recomputed them on 2026-09-13 (they were rounded estimates before); the aggregate row is still the paper's rounded estimate, dated 2026-04-21. Nothing on this site re-derives them. Before each deploy they are compared with the copy monogate.org shows, so a change there fails the next deploy until this table matches.

Machine-Checked Theorems (Lean 4)

The cost theory's named results, mechanised in MachLib/CostTheory.lean ↗ — pure Nat, #print axioms shows only propext / Quot.sound, no sorryAx, no MachLib field axioms. Status is honest: the empirical mean-cost ordering (C < B < A < D) and the T42-QCC conjecture are not claimed as proven.

ResultStatementLean theoremStatus
T38-NNPNo-Nesting Penalty — composing operators is additive, no adapter/depth overheadno_nesting_penaltymachine-checked
T38Cost = Naive − Pattern − Sharing — the two savings account for exactly the naive/actual gapt38_decompositionmachine-checked
T40Additive Cost Law — independent branches sum exactly; sums are association-independentadditive_cost_law · sum_assoc_invariantmachine-checked
T41-ISOCross-domain cost invariance — isomorphic DAG topologies cost the sameiso_costmachine-checked
T42O(N) single-sum law — N equal-cost terms cost (α₀+3)·N − 3cost_flatSummachine-checked
T42²O(N²) double-sum law — a nested N×N sum is an exact quadraticcost_doubleSummachine-checked
P1–P3Basic properties — non-negativity, terminal characterisation, subadditivityp1_nonneg · p2_*_cost_ge · p3_subadditive_*machine-checked
T41Four structural classes — classification well-defined; rational is the cost floor, mixed the ceilingclassify · rational_floor · mixed_ceilingstructural core
P4Algebraic invariance — cost as a function of the computed FUNCTION, not the treeout of scope
T42-QCCQuadratic Ceiling Conjecture — no closed form exceeds O(N²); empirical over 187 equationsconjecture

Key results

  • Core table, recounted in F16 (2026-09-13): 15n / 79.5% vs 73n naive on the positive domain and 18n / 66.7% vs 54n naive on the 6-op general basket; 14n / 80.8% when ln counts as one EXL node. Best-known upper bounds, evaluated numerically in the monogate repository; the totals are checked against the monogate.superbest library before each deploy. Extended operators do not change these counts
  • LSE corrected: ln(e^x+e^y) = 4n in F16 (was 5n); 2n in 23-op via EEA+ln
  • Taxonomy: 23 operators catalogued. CONJ_NO_OP_24 (no 24th operator) is a conjecture, argued on paper; no Lean proof exists
  • softplus = 2n in both layers via EML(x,1/e)+ln (Category B: genuine F16)
  • Aggregate genuine F16 savings: ~12% across the taxonomy lock's catalogs; ~40% with 23-op extended (rounded estimates dated 2026-04-21, not re-derived; compared with monogate.org's copy before each deploy)

Research: monogate.org/superbest ↗