Knowledge

Large deviations of Gaussian random functions

Source đź“ť

990: 542:
to the sphere). The derivative is random, with zero expectation and some standard deviation. The latter may depend on the point and the direction. However, if it does not depend, then it is equal to
1423:
The basic statement given above is a simple special case of a much more general (and difficult) theory stated by Adler. For a detailed presentation of this special case see Tsirelson's lectures.
342: 607: 1356: 885: 685: 1103: 1243: 1161: 1040: 794: 1391: 1274: 849: 759: 262: 1320: 1211: 1129: 1070: 397: 227: 362: 1294: 1185: 814: 716: 650: 627: 536: 512: 489: 465: 441: 421: 201: 181: 161: 141: 121: 101: 893: 1456: 49: 1163:
is non-empty; its Euler characteristic may take various values, depending on the topology of the set (the number of
267: 37: 33: 75:. Though rare in a small domain (of space or/and time), large deviations may be quite usual in a large domain. 1498: 545: 400: 56: 996: 719: 1000: 1325: 854: 655: 1412: 1075: 1216: 1134: 1013: 767: 1361: 1252: 819: 737: 515: 232: 1407: 1480: 1299: 1190: 1108: 1049: 367: 206: 688: 444: 347: 8: 1279: 1249:(which is easy to guess, but quite difficult to prove). Thus, its Euler characteristic 1170: 799: 701: 635: 612: 521: 497: 474: 450: 426: 406: 186: 166: 146: 126: 106: 86: 68: 1452: 1004: 762: 1402: 1451:
Robert J. Adler, Jonathan E. Taylor, "Random fields and geometry", Springer 2007.
1438:
Robert J. Adler, "On excursion sets, tube formulas and maxima of random fields",
1164: 468: 64: 21: 1492: 1439: 723: 514:
is easy to determine in the important special case described in terms of the
25: 539: 1467:
Robert J. Adler, "Some new random field tools for spatial analysis",
1043: 123:
on the (two-dimensional) sphere. Assume that the expected value of
63:
Sometimes, a value of a Gaussian random function deviates from its
1468: 1440:
The Annals of Applied Probability 2000, Vol. 10, No. 1, 1–74
1246: 734:
The clue to the theory sketched above is, Euler characteristic
41: 726:), and reaches its maximum at a single point (almost surely). 163:(at every point of the sphere), and the standard deviation of 692: 985:{\displaystyle E(\chi _{a})=Ca\exp(-a^{2}/2)+2P(\xi >a)} 1167:, and possible holes in these components). However, if 538:
at a given point (of the sphere) in a given direction (
729: 1364: 1328: 1302: 1282: 1255: 1219: 1193: 1173: 1137: 1111: 1078: 1052: 1016: 896: 857: 822: 802: 770: 740: 704: 658: 638: 615: 548: 524: 500: 477: 453: 429: 409: 370: 350: 270: 235: 209: 189: 169: 149: 129: 109: 89: 471:
of the approximation decays exponentially for large
103:be the maximal value of a Gaussian random function 1385: 1350: 1314: 1288: 1268: 1237: 1205: 1179: 1155: 1123: 1097: 1064: 1034: 984: 879: 851:. Its expected value (in other words, mean value) 843: 808: 788: 753: 710: 679: 644: 621: 601: 530: 506: 483: 459: 435: 415: 391: 356: 336: 256: 221: 195: 175: 155: 135: 115: 95: 1490: 203:(at every point of the sphere). Then, for large 995:(which is far from being trivial, and involves 1245:is usually a small, slightly deformed disk or 337:{\displaystyle Ca\exp(-a^{2}/2)+2P(\xi >a)} 1232: 1220: 1150: 1138: 1029: 1017: 783: 771: 1434: 1432: 1430: 1461: 1427: 44:are useful (for example) when analysing 1491: 1445: 602:{\displaystyle (\pi /2)^{1/4}C^{1/2}} 423:is a constant; it does not depend on 50:cosmic microwave background radiation 730:The clue: mean Euler characteristic 13: 1418: 78: 20:– of either one variable (a 14: 1510: 40:. Gaussian random fields on the 38:multivariate normal distribution 34:finite-dimensional distribution 24:), or two or more variables (a 1474: 1380: 1368: 1345: 1332: 979: 967: 955: 931: 913: 900: 887:can be calculated explicitly: 874: 861: 832: 826: 674: 662: 564: 549: 386: 374: 331: 319: 307: 283: 251: 239: 1: 1351:{\displaystyle E(\chi _{a})} 880:{\displaystyle E(\chi _{a})} 680:{\displaystyle P(\xi >a)} 401:standard normal distribution 57:positron emission tomography 7: 1396: 1098:{\displaystyle \chi _{a}=0} 720:continuously differentiable 59:(see, pp. 9–10). 10: 1515: 1442:. (Special invited paper.) 1238:{\displaystyle \{X>a\}} 1156:{\displaystyle \{X>a\}} 1105:. In the other case, when 1035:{\displaystyle \{X>a\}} 816:(of the sphere) such that 789:{\displaystyle \{X>a\}} 609:(for the sphere of radius 52:(see, pp. 8–9); 1386:{\displaystyle P(M>a)} 1269:{\displaystyle \chi _{a}} 844:{\displaystyle X(t)>a} 754:{\displaystyle \chi _{a}} 257:{\displaystyle P(M>a)} 55:brain images obtained by 1481:Lectures of B. Tsirelson 1413:Large deviations theory 691:of the sphere (for the 1483:(especially, Sect. 5). 1387: 1352: 1316: 1315:{\displaystyle M>a} 1290: 1270: 1239: 1207: 1206:{\displaystyle M>a} 1181: 1157: 1125: 1124:{\displaystyle M>a} 1099: 1066: 1065:{\displaystyle M<a} 1036: 986: 881: 845: 810: 790: 755: 712: 681: 646: 623: 603: 532: 516:directional derivative 508: 485: 461: 437: 417: 393: 392:{\displaystyle N(0,1)} 358: 338: 258: 223: 222:{\displaystyle a>0} 197: 177: 157: 137: 117: 97: 1408:Gaussian random field 1388: 1353: 1317: 1291: 1271: 1240: 1208: 1182: 1158: 1126: 1100: 1067: 1037: 997:Poincaré–Hopf theorem 987: 882: 846: 811: 791: 756: 713: 682: 647: 624: 604: 533: 509: 486: 462: 443:, but depends on the 438: 418: 394: 359: 339: 259: 224: 198: 178: 158: 138: 118: 98: 48:the anomalies in the 1499:Stochastic processes 1362: 1326: 1300: 1280: 1276:is usually equal to 1253: 1217: 1191: 1171: 1165:connected components 1135: 1109: 1076: 1050: 1014: 1001:Gauss–Bonnet theorem 894: 855: 820: 800: 768: 738: 702: 689:Euler characteristic 656: 636: 613: 546: 522: 498: 475: 451: 445:correlation function 427: 407: 368: 357:{\displaystyle \xi } 348: 268: 233: 207: 187: 167: 147: 127: 107: 87: 28:) – is called 698:It is assumed that 69:standard deviations 1383: 1348: 1312: 1286: 1266: 1235: 1203: 1177: 1153: 1121: 1095: 1062: 1032: 982: 877: 841: 806: 786: 751: 708: 677: 642: 619: 599: 528: 504: 481: 457: 433: 413: 389: 354: 334: 254: 219: 193: 173: 153: 133: 113: 93: 1457:978-0-387-48112-8 1289:{\displaystyle 1} 1180:{\displaystyle a} 809:{\displaystyle t} 711:{\displaystyle X} 645:{\displaystyle 2} 622:{\displaystyle 1} 531:{\displaystyle X} 507:{\displaystyle C} 484:{\displaystyle a} 467:(see below). The 460:{\displaystyle X} 436:{\displaystyle a} 416:{\displaystyle C} 196:{\displaystyle 1} 176:{\displaystyle X} 156:{\displaystyle 0} 136:{\displaystyle X} 116:{\displaystyle X} 96:{\displaystyle M} 1506: 1484: 1478: 1472: 1465: 1459: 1449: 1443: 1436: 1403:Gaussian process 1392: 1390: 1389: 1384: 1357: 1355: 1354: 1349: 1344: 1343: 1321: 1319: 1318: 1313: 1295: 1293: 1292: 1287: 1275: 1273: 1272: 1267: 1265: 1264: 1244: 1242: 1241: 1236: 1212: 1210: 1209: 1204: 1186: 1184: 1183: 1178: 1162: 1160: 1159: 1154: 1130: 1128: 1127: 1122: 1104: 1102: 1101: 1096: 1088: 1087: 1071: 1069: 1068: 1063: 1041: 1039: 1038: 1033: 991: 989: 988: 983: 951: 946: 945: 912: 911: 886: 884: 883: 878: 873: 872: 850: 848: 847: 842: 815: 813: 812: 807: 795: 793: 792: 787: 760: 758: 757: 752: 750: 749: 717: 715: 714: 709: 686: 684: 683: 678: 651: 649: 648: 643: 632:The coefficient 628: 626: 625: 620: 608: 606: 605: 600: 598: 597: 593: 580: 579: 575: 559: 537: 535: 534: 529: 513: 511: 510: 505: 490: 488: 487: 482: 466: 464: 463: 458: 442: 440: 439: 434: 422: 420: 419: 414: 398: 396: 395: 390: 363: 361: 360: 355: 343: 341: 340: 335: 303: 298: 297: 263: 261: 260: 255: 228: 226: 225: 220: 202: 200: 199: 194: 182: 180: 179: 174: 162: 160: 159: 154: 142: 140: 139: 134: 122: 120: 119: 114: 102: 100: 99: 94: 1514: 1513: 1509: 1508: 1507: 1505: 1504: 1503: 1489: 1488: 1487: 1479: 1475: 1469:arXiv:0805.1031 1466: 1462: 1450: 1446: 1437: 1428: 1421: 1419:Further reading 1399: 1363: 1360: 1359: 1339: 1335: 1327: 1324: 1323: 1322:). This is why 1301: 1298: 1297: 1281: 1278: 1277: 1260: 1256: 1254: 1251: 1250: 1218: 1215: 1214: 1192: 1189: 1188: 1172: 1169: 1168: 1136: 1133: 1132: 1110: 1107: 1106: 1083: 1079: 1077: 1074: 1073: 1072:; in this case 1051: 1048: 1047: 1015: 1012: 1011: 947: 941: 937: 907: 903: 895: 892: 891: 868: 864: 856: 853: 852: 821: 818: 817: 801: 798: 797: 769: 766: 765: 745: 741: 739: 736: 735: 732: 703: 700: 699: 687:is in fact the 657: 654: 653: 637: 634: 633: 614: 611: 610: 589: 585: 581: 571: 567: 563: 555: 547: 544: 543: 523: 520: 519: 499: 496: 495: 476: 473: 472: 452: 449: 448: 428: 425: 424: 408: 405: 404: 369: 366: 365: 364:is distributed 349: 346: 345: 299: 293: 289: 269: 266: 265: 234: 231: 230: 208: 205: 204: 188: 185: 184: 168: 165: 164: 148: 145: 144: 128: 125: 124: 108: 105: 104: 88: 85: 84: 81: 79:Basic statement 73:large deviation 18:random function 12: 11: 5: 1512: 1502: 1501: 1486: 1485: 1473: 1460: 1444: 1425: 1420: 1417: 1416: 1415: 1410: 1405: 1398: 1395: 1382: 1379: 1376: 1373: 1370: 1367: 1347: 1342: 1338: 1334: 1331: 1311: 1308: 1305: 1285: 1263: 1259: 1234: 1231: 1228: 1225: 1222: 1202: 1199: 1196: 1176: 1152: 1149: 1146: 1143: 1140: 1120: 1117: 1114: 1094: 1091: 1086: 1082: 1061: 1058: 1055: 1031: 1028: 1025: 1022: 1019: 1005:Rice's formula 993: 992: 981: 978: 975: 972: 969: 966: 963: 960: 957: 954: 950: 944: 940: 936: 933: 930: 927: 924: 921: 918: 915: 910: 906: 902: 899: 876: 871: 867: 863: 860: 840: 837: 834: 831: 828: 825: 805: 796:of all points 785: 782: 779: 776: 773: 748: 744: 731: 728: 707: 695:it vanishes). 676: 673: 670: 667: 664: 661: 641: 618: 596: 592: 588: 584: 578: 574: 570: 566: 562: 558: 554: 551: 527: 503: 480: 469:relative error 456: 432: 412: 388: 385: 382: 379: 376: 373: 353: 333: 330: 327: 324: 321: 318: 315: 312: 309: 306: 302: 296: 292: 288: 285: 282: 279: 276: 273: 253: 250: 247: 244: 241: 238: 218: 215: 212: 192: 172: 152: 132: 112: 92: 80: 77: 65:expected value 61: 60: 53: 22:random process 9: 6: 4: 3: 2: 1511: 1500: 1497: 1496: 1494: 1482: 1477: 1470: 1464: 1458: 1454: 1448: 1441: 1435: 1433: 1431: 1426: 1424: 1414: 1411: 1409: 1406: 1404: 1401: 1400: 1394: 1377: 1374: 1371: 1365: 1340: 1336: 1329: 1309: 1306: 1303: 1283: 1261: 1257: 1248: 1229: 1226: 1223: 1213:then the set 1200: 1197: 1194: 1187:is large and 1174: 1166: 1147: 1144: 1141: 1118: 1115: 1112: 1092: 1089: 1084: 1080: 1059: 1056: 1053: 1045: 1026: 1023: 1020: 1008: 1006: 1002: 998: 976: 973: 970: 964: 961: 958: 952: 948: 942: 938: 934: 928: 925: 922: 919: 916: 908: 904: 897: 890: 889: 888: 869: 865: 858: 838: 835: 829: 823: 803: 780: 777: 774: 764: 746: 742: 727: 725: 724:almost surely 721: 705: 696: 694: 690: 671: 668: 665: 659: 639: 630: 616: 594: 590: 586: 582: 576: 572: 568: 560: 556: 552: 541: 525: 517: 501: 494:The constant 492: 478: 470: 454: 446: 430: 410: 402: 383: 380: 377: 371: 351: 328: 325: 322: 316: 313: 310: 304: 300: 294: 290: 286: 280: 277: 274: 271: 248: 245: 242: 236: 216: 213: 210: 190: 170: 150: 130: 110: 90: 76: 74: 70: 66: 58: 54: 51: 47: 46: 45: 43: 39: 35: 31: 27: 23: 19: 1476: 1463: 1447: 1422: 1358:is close to 1296:(given that 1009: 994: 733: 697: 631: 493: 264:is close to 82: 72: 71:. This is a 62: 29: 26:random field 17: 15: 1131:, the set 67:by several 540:tangential 1337:χ 1258:χ 1081:χ 1046:whenever 1044:empty set 971:ξ 935:− 929:⁡ 905:χ 866:χ 743:χ 718:is twice 666:ξ 553:π 352:ξ 323:ξ 287:− 281:⁡ 32:if every 1493:Category 1397:See also 1010:The set 344:, where 30:Gaussian 1247:ellipse 1042:is the 1007:etc.). 761:of the 652:before 403:), and 1455:  42:sphere 693:torus 399:(the 36:is a 1453:ISBN 1375:> 1307:> 1227:> 1198:> 1145:> 1116:> 1057:< 1024:> 974:> 836:> 778:> 669:> 326:> 246:> 214:> 83:Let 926:exp 763:set 629:). 518:of 447:of 278:exp 183:is 143:is 1495:: 1429:^ 1393:. 1003:, 999:, 491:. 229:, 16:A 1471:. 1381:) 1378:a 1372:M 1369:( 1366:P 1346:) 1341:a 1333:( 1330:E 1310:a 1304:M 1284:1 1262:a 1233:} 1230:a 1224:X 1221:{ 1201:a 1195:M 1175:a 1151:} 1148:a 1142:X 1139:{ 1119:a 1113:M 1093:0 1090:= 1085:a 1060:a 1054:M 1030:} 1027:a 1021:X 1018:{ 980:) 977:a 968:( 965:P 962:2 959:+ 956:) 953:2 949:/ 943:2 939:a 932:( 923:a 920:C 917:= 914:) 909:a 901:( 898:E 875:) 870:a 862:( 859:E 839:a 833:) 830:t 827:( 824:X 804:t 784:} 781:a 775:X 772:{ 747:a 722:( 706:X 675:) 672:a 663:( 660:P 640:2 617:1 595:2 591:/ 587:1 583:C 577:4 573:/ 569:1 565:) 561:2 557:/ 550:( 526:X 502:C 479:a 455:X 431:a 411:C 387:) 384:1 381:, 378:0 375:( 372:N 332:) 329:a 320:( 317:P 314:2 311:+ 308:) 305:2 301:/ 295:2 291:a 284:( 275:a 272:C 252:) 249:a 243:M 240:( 237:P 217:0 211:a 191:1 171:X 151:0 131:X 111:X 91:M

Index

random process
random field
finite-dimensional distribution
multivariate normal distribution
sphere
cosmic microwave background radiation
positron emission tomography
expected value
standard deviations
standard normal distribution
correlation function
relative error
directional derivative
tangential
Euler characteristic
torus
continuously differentiable
almost surely
set
Poincaré–Hopf theorem
Gauss–Bonnet theorem
Rice's formula
empty set
connected components
ellipse
Gaussian process
Gaussian random field
Large deviations theory

Text is available under the Creative Commons Attribution-ShareAlike License. Additional terms may apply.

↑