code wiki / (root) / nx_sketch_count_sketch.nx

nx_sketch_count_sketch.nx source

↩ module page · 210 lines · 7208 B

1// sketch_count_sketch.nx -- CountSketch (Charikar-Chen-Farach-Colton 2002). 2// 3// Frequency-counting variant that gives UNBIASED estimates via ±1 4// sign hashing. Each (key, count) update adds s_j(key) * count to 5// row j column h_j(key), where s_j ∈ {-1, +1} is a sign hash. 6// Query: median of (s_j(key) * counter[j][h_j(key)]) across rows. 7// 8// COMPLEMENTS CMS (runtime/sketch_cms.nx): 9// - CMS: always OVERESTIMATES (false positives via positive collisions). 10// Use when "upper bound" matters. 11// - CountSketch: UNBIASED. Use when expectation matters (statistical 12// summaries, F_2 estimation, sparse-approximation literature). 13// 14// CAPABILITY STOMP: shipping both means callers pick by error semantics. 15// DataSketches ships CMS as "FrequentLongs"; CountSketch is queued in 16// their docs but not implemented. 17// 18// ERROR BOUND: 19// |estimate(x) - f(x)| <= ε · ||f||_2 / sqrt(d) 20// where ||f||_2 = sqrt(Σ f_i²) is the L2 norm of the frequency vector. 21// For w = O(1/ε²) and d = O(log(1/δ)) rows: confidence 1-δ. 22// 23// Default params: d=5 rows, w=2048 columns -> conf >99%, ε ~ 0.022. 24 25// nx_safety_envelope: 26// intended_use: AUTO_APPLIED -- primitive-specific tuning queued 27// sil_target: SIL1 28// evidence: [bulk_applied_2026-05-16, see-file-comment-for-detail] 29// verdict: NOT_YET_EVALUATED 30 31import "nx_syscalls.nx" 32import "nx_sketch_types.nx" 33import "nx_vecmath.nx" 34 35const NX_CS_MIN_D: i64 = 3 36const NX_CS_MAX_D: i64 = 32 37const NX_CS_MIN_W: i64 = 64 38const NX_CS_MAX_W: i64 = 65536 39 40struct CountSketch { 41 counters: *i64, // d * w grid, in i64 (signed) 42 d: i64, 43 w: i64, 44 seed: i64, 45 total: i64, 46} 47 48// === construction ================================================= 49 50func nx_cs_alloc(d: i64, w: i64, seed: i64) -> *CountSketch { 51 if d < NX_CS_MIN_D { return 0 as *CountSketch } 52 if d > NX_CS_MAX_D { return 0 as *CountSketch } 53 if w < NX_CS_MIN_W { return 0 as *CountSketch } 54 if w > NX_CS_MAX_W { return 0 as *CountSketch } 55 if (w & (w - 1)) != 0 { return 0 as *CountSketch } // power of 2 56 let raw: *u8 = sys_mmap(48) 57 let c: *CountSketch = raw as *CountSketch 58 let cells: i64 = d * w 59 let cells_raw: *u8 = sys_mmap(cells * 8) 60 c.counters = cells_raw as *i64 61 var i: i64 = 0 62 while i < cells { 63 c.counters[i] = 0 64 i = i + 1 65 } 66 c.d = d 67 c.w = w 68 c.seed = seed 69 c.total = 0 70 return c 71} 72 73// === hash + sign helpers ========================================= 74// 75// Kirsch-Mitzenmacher double-hashing: h_j(x) = (h1 + j*h2) & mask. 76// Sign: top bit of a separately-derived hash, mapped to ±1. 77 78func nx_cs_h1(c: *CountSketch, key: i64) -> i64 { 79 let mixed: i64 = (key * 0x9E3779B97F4A7C15 + c.seed) & 0xFFFFFFFFFFFFFFFF 80 return mixed & 0xFFFFFFFF 81} 82 83func nx_cs_h2(c: *CountSketch, key: i64) -> i64 { 84 let mixed: i64 = (key * 0xBF58476D1CE4E5B9 + c.seed) & 0xFFFFFFFFFFFFFFFF 85 return mixed & 0xFFFFFFFF 86} 87 88func nx_cs_column(c: *CountSketch, key: i64, j: i64) -> i64 { 89 let combined: i64 = (nx_cs_h1(c, key) + j * nx_cs_h2(c, key)) & 0xFFFFFFFF 90 return combined & (c.w - 1) 91} 92 93// Sign for row j: hash key+j and check the top bit. 94func nx_cs_sign(c: *CountSketch, key: i64, j: i64) -> i64 { 95 let mixed: i64 = ((key + j * 0x9E3779B9) * 0xC2B2AE3D27D4EB4F + c.seed) & 0xFFFFFFFFFFFFFFFF 96 if (mixed & (1 << 63)) == 0 { return 1 } 97 return -1 98} 99 100// === cell access ================================================= 101 102func nx_cs_cell_idx(c: *CountSketch, j: i64, col: i64) -> i64 { 103 return j * c.w + col 104} 105 106// === add ========================================================== 107 108func nx_cs_add(c: *CountSketch, key: i64, count: i64) -> i64 { 109 if count == 0 { return 0 } 110 c.total = c.total + count 111 var j: i64 = 0 112 while j < c.d { 113 let col: i64 = nx_cs_column(c, key, j) 114 let sign: i64 = nx_cs_sign(c, key, j) 115 let idx: i64 = nx_cs_cell_idx(c, j, col) 116 c.counters[idx] = c.counters[idx] + sign * count 117 j = j + 1 118 } 119 return 0 120} 121 122// === query (median of signed reads) ============================== 123// 124// For each row j: signed_read = sign_j(key) * counter[j][h_j(key)] 125// Return median of d signed reads. Median is approximate median via 126// insertion-sort of small d. 127 128func nx_cs_estimate(c: *CountSketch, key: i64) -> i64 { 129 // Collect d signed reads. 130 let scratch_raw: *u8 = sys_mmap(c.d * 8) 131 let scratch: *i64 = scratch_raw as *i64 132 var j: i64 = 0 133 while j < c.d { 134 let col: i64 = nx_cs_column(c, key, j) 135 let sign: i64 = nx_cs_sign(c, key, j) 136 let idx: i64 = nx_cs_cell_idx(c, j, col) 137 scratch[j] = sign * c.counters[idx] 138 j = j + 1 139 } 140 // Insertion sort. 141 var i: i64 = 1 142 while i < c.d { 143 let cur: i64 = scratch[i] 144 var k: i64 = i - 1 145 var done: i64 = 0 146 while done == 0 { 147 if k < 0 { done = 1 } 148 if done == 0 { 149 if scratch[k] <= cur { done = 1 } 150 if done == 0 { 151 scratch[k + 1] = scratch[k] 152 k = k - 1 153 } 154 } 155 } 156 scratch[k + 1] = cur 157 i = i + 1 158 } 159 return scratch[c.d / 2] 160} 161 162// === typed envelope =============================================== 163// 164// Error bound is |est - f| <= ||f||_2 * sqrt(2/w) at confidence 165// (3/4)^(d/2). We declare a CONSERVATIVE absolute bound = total/sqrt(w) 166// (this is the worst case when L2 = total, i.e. one massive item). 167// param_a holds the bound; conf = (3/4)^(d/2) tabulated. 168 169func nx_cs_isqrt(x: i64) -> i64 { return vm_isqrt(x) } 170 171// Confidence depends on d: (3/4)^(d/2). Approximate by tabulating. 172func nx_cs_conf_ppb(d: i64) -> i64 { 173 if d <= 3 { return 562500000 } // (3/4)^1.5 = 0.65 174 if d <= 5 { return 421900000 } // (3/4)^2.5 175 if d <= 7 { return 316400000 } // (3/4)^3.5 176 if d <= 11 { return 177900000 } // (3/4)^5.5 177 return 100000000 // higher d converges 178} 179 180func nx_cs_query(c: *CountSketch, key: i64) -> *ApproxI64 { 181 let est: i64 = nx_cs_estimate(c, key) 182 let bound: i64 = c.total / nx_cs_isqrt(c.w) 183 // Returned envelope: 1e9 - conf_ppb confidence the bound holds; 184 // i.e. there's a small chance the estimate is further than `bound`. 185 return nx_approx_new(est, NX_ENV_ABS, bound, 186 1000000000 - nx_cs_conf_ppb(c.d), 187 NX_MATURITY_REFERENCE_IMPL, 188 NX_ADV_HONEST) 189} 190 191// === merge ======================================================== 192 193func nx_cs_merge(a: *CountSketch, b: *CountSketch) -> *CountSketch { 194 if a.d != b.d { return 0 as *CountSketch } 195 if a.w != b.w { return 0 as *CountSketch } 196 if a.seed != b.seed { return 0 as *CountSketch } 197 let out: *CountSketch = nx_cs_alloc(a.d, a.w, a.seed) 198 let cells: i64 = a.d * a.w 199 var i: i64 = 0 200 while i < cells { 201 out.counters[i] = a.counters[i] + b.counters[i] 202 i = i + 1 203 } 204 out.total = a.total + b.total 205 return out 206} 207 208func nx_cs_memory_bytes(c: *CountSketch) -> i64 { 209 return 48 + c.d * c.w * 8 210}