File size: 157,448 Bytes
077605d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636 3637 3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648 3649 3650 3651 3652 3653 3654 3655 3656 3657 3658 3659 3660 3661 3662 3663 3664 3665 3666 3667 3668 3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713 3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748 3749 3750 3751 3752 3753 3754 3755 3756 3757 3758 3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826 3827 3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848 3849 3850 3851 3852 3853 3854 3855 3856 3857 3858 3859 3860 3861 3862 3863 3864 3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892 3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951 3952 3953 3954 3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982 3983 3984 3985 3986 3987 3988 3989 3990 3991 3992 3993 3994 3995 3996 3997 3998 3999 4000 4001 4002 4003 4004 4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031 4032 4033 4034 4035 4036 4037 4038 4039 4040 4041 4042 4043 4044 4045 4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057 4058 4059 4060 4061 4062 4063 4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077 4078 4079 4080 4081 4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095 4096 4097 4098 4099 4100 4101 4102 4103 4104 4105 4106 4107 4108 4109 4110 4111 4112 4113 4114 4115 4116 4117 4118 4119 4120 4121 4122 4123 4124 4125 4126 4127 4128 4129 4130 4131 4132 4133 4134 4135 4136 4137 4138 4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167 4168 4169 4170 4171 4172 4173 4174 4175 4176 4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190 4191 4192 4193 4194 4195 4196 4197 4198 4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238 4239 4240 4241 4242 4243 4244 4245 4246 4247 4248 4249 4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261 4262 4263 4264 4265 4266 4267 4268 4269 4270 4271 4272 4273 4274 4275 4276 4277 4278 4279 4280 4281 4282 4283 4284 4285 4286 4287 4288 4289 4290 4291 4292 4293 4294 4295 4296 4297 4298 4299 4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312 4313 4314 4315 4316 4317 4318 4319 4320 4321 4322 4323 4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429 4430 4431 4432 4433 4434 4435 4436 4437 4438 4439 4440 4441 4442 4443 4444 4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458 4459 4460 4461 4462 4463 4464 4465 4466 4467 4468 4469 4470 4471 4472 4473 4474 4475 4476 4477 4478 4479 4480 4481 4482 4483 4484 4485 4486 4487 4488 4489 4490 4491 4492 4493 4494 4495 4496 4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525 4526 4527 4528 4529 4530 4531 4532 4533 4534 4535 4536 4537 4538 4539 4540 4541 4542 4543 4544 4545 4546 4547 4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576 4577 4578 4579 4580 4581 4582 4583 4584 4585 4586 4587 4588 4589 4590 4591 4592 4593 4594 4595 4596 4597 4598 4599 4600 4601 4602 4603 4604 4605 4606 4607 4608 4609 4610 4611 4612 4613 4614 4615 4616 4617 4618 4619 4620 4621 4622 4623 4624 4625 4626 4627 4628 4629 4630 4631 4632 4633 4634 4635 4636 4637 4638 4639 4640 4641 4642 4643 4644 4645 4646 4647 4648 4649 4650 4651 4652 4653 4654 4655 4656 4657 4658 4659 4660 4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675 4676 4677 4678 4679 4680 4681 4682 4683 4684 4685 4686 4687 4688 4689 4690 4691 4692 4693 4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706 4707 4708 4709 4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722 4723 4724 4725 4726 4727 4728 4729 4730 4731 4732 4733 4734 4735 4736 4737 4738 4739 4740 4741 4742 4743 4744 4745 4746 4747 4748 4749 4750 4751 4752 4753 4754 4755 4756 4757 4758 4759 4760 4761 4762 4763 4764 4765 4766 4767 4768 4769 4770 4771 4772 4773 4774 4775 4776 4777 4778 4779 4780 4781 4782 4783 4784 4785 4786 4787 4788 4789 4790 4791 4792 4793 4794 4795 4796 4797 4798 4799 4800 4801 4802 4803 4804 4805 4806 4807 4808 4809 4810 4811 4812 4813 4814 4815 4816 4817 4818 4819 4820 4821 4822 4823 4824 4825 4826 4827 4828 4829 4830 4831 4832 4833 4834 4835 4836 4837 4838 4839 4840 4841 4842 4843 4844 4845 4846 4847 4848 4849 4850 4851 4852 4853 4854 4855 4856 4857 4858 4859 4860 4861 4862 4863 4864 4865 4866 4867 4868 4869 4870 4871 4872 4873 4874 4875 4876 4877 4878 4879 4880 4881 4882 4883 4884 4885 4886 4887 4888 4889 4890 4891 4892 4893 4894 4895 4896 4897 4898 4899 4900 4901 4902 4903 4904 4905 4906 4907 4908 4909 4910 4911 4912 4913 4914 4915 4916 4917 4918 4919 4920 4921 4922 4923 4924 4925 4926 4927 4928 4929 4930 4931 4932 4933 4934 4935 4936 4937 4938 4939 4940 4941 4942 4943 4944 4945 4946 4947 4948 4949 4950 4951 4952 4953 4954 4955 4956 4957 4958 4959 4960 4961 4962 4963 4964 4965 4966 4967 4968 4969 4970 4971 4972 4973 4974 4975 4976 4977 4978 4979 4980 4981 4982 4983 4984 4985 4986 4987 4988 4989 4990 4991 4992 4993 4994 4995 4996 4997 4998 4999 5000 5001 5002 5003 5004 5005 5006 5007 5008 5009 5010 5011 5012 5013 5014 5015 5016 5017 5018 5019 5020 5021 5022 5023 5024 5025 5026 5027 5028 5029 5030 5031 5032 5033 5034 5035 5036 5037 5038 5039 5040 5041 5042 5043 5044 5045 5046 5047 5048 5049 5050 5051 5052 5053 5054 5055 5056 5057 5058 5059 5060 5061 5062 5063 5064 5065 5066 5067 5068 5069 5070 5071 5072 5073 5074 5075 5076 5077 5078 5079 5080 5081 5082 5083 5084 5085 5086 5087 5088 5089 5090 5091 5092 5093 5094 5095 5096 5097 5098 5099 5100 5101 5102 5103 5104 5105 5106 5107 5108 5109 5110 5111 5112 5113 5114 5115 5116 5117 5118 5119 5120 5121 5122 5123 5124 5125 5126 5127 5128 5129 5130 5131 5132 5133 5134 5135 5136 5137 5138 5139 5140 5141 5142 5143 5144 5145 5146 5147 5148 5149 5150 5151 5152 5153 5154 5155 5156 5157 5158 5159 5160 5161 5162 5163 5164 5165 5166 5167 5168 5169 5170 5171 5172 5173 5174 5175 5176 5177 5178 5179 5180 5181 5182 5183 5184 5185 5186 5187 5188 5189 5190 5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201 5202 5203 5204 5205 5206 5207 5208 5209 5210 5211 5212 5213 5214 5215 5216 5217 5218 5219 5220 5221 5222 5223 5224 5225 5226 5227 5228 5229 5230 5231 5232 5233 5234 5235 5236 5237 5238 5239 5240 5241 5242 5243 5244 5245 5246 5247 5248 5249 5250 5251 5252 5253 5254 5255 5256 5257 5258 5259 5260 5261 5262 5263 5264 5265 5266 5267 5268 5269 5270 5271 5272 5273 5274 5275 5276 5277 5278 5279 5280 5281 5282 5283 5284 5285 5286 5287 5288 5289 5290 5291 5292 5293 5294 5295 5296 5297 5298 5299 5300 5301 5302 5303 5304 5305 5306 5307 5308 5309 5310 5311 5312 5313 5314 5315 5316 5317 5318 5319 5320 5321 5322 5323 5324 5325 5326 5327 5328 5329 5330 5331 5332 5333 5334 5335 5336 5337 5338 5339 5340 5341 5342 5343 5344 5345 5346 5347 5348 5349 5350 5351 5352 5353 5354 5355 5356 5357 5358 5359 5360 5361 5362 5363 5364 5365 5366 5367 5368 5369 5370 5371 5372 5373 5374 5375 5376 5377 5378 5379 5380 5381 5382 5383 5384 5385 5386 5387 5388 5389 5390 5391 5392 5393 5394 5395 5396 5397 5398 5399 5400 5401 5402 5403 5404 5405 5406 5407 5408 5409 5410 5411 5412 5413 5414 5415 5416 5417 5418 5419 5420 5421 5422 5423 5424 5425 5426 5427 5428 5429 5430 5431 5432 5433 5434 5435 5436 5437 5438 5439 5440 5441 5442 5443 5444 5445 5446 5447 5448 5449 5450 5451 5452 5453 5454 5455 5456 5457 5458 5459 5460 5461 5462 5463 5464 5465 5466 5467 5468 5469 5470 5471 5472 5473 5474 5475 5476 5477 5478 5479 5480 5481 5482 5483 5484 5485 5486 5487 5488 5489 5490 5491 5492 5493 5494 5495 5496 5497 5498 5499 5500 5501 5502 5503 5504 5505 5506 5507 5508 5509 5510 5511 | Transformers, parallel computation, and logarithmic depth
Clayton Sanford1, Daniel Hsu1, and Matus Telgarsky2
1Department of Computer Science, Columbia University, New York, NY, USA
2Courant Institute, New York University, New York, NY, USA
February 15, 2024
Abstract
We show that a constant number of self-attention layers can efficiently simulate—and be simulated
by—a constant number of communication rounds of Massively Parallel Computation. As a consequence,
we show that logarithmic depth is sufficient for transformers to solve basic computational tasks that
cannot be efficiently solved by several other neural sequence models and sub-quadratic transformer
approximations. We thus establish parallelism as a key distinguishing property of transformers.
1
Introduction
The transformer (Vaswani et al., 2017) has emerged as the dominant neural architecture for many sequential
modeling tasks such as machine translation (Radford et al., 2019) and protein folding (Jumper et al., 2021).
Reasons for the success of transformers include suitability to modern hardware and training stability: unlike
in recurrent models, inference and training can be efficiently parallelized, and training is less vulnerable to
vanishing and exploding gradients. However, the advantages of transformers over other neural architectures can
be understood more fundamentally via the lens of representation, which regards neural nets as parameterized
functions and asks what they can efficiently compute.
Many previous theoretical studies of transformers establish (approximation-theoretic and computational)
universality properties, but only at large model sizes (Yun et al., 2020; Pérez et al., 2021). These results
are not unique to transformers and reveal little about which tasks can be solved in a size-efficient manner.
Several other works (e.g., Hahn, 2020; Merrill and Sabharwal, 2022; Sanford et al., 2023) give fine-grained
representational results in the scaling regime where context length grows but model depth is constant. In this
regime, basic algorithmic tasks like matching parentheses and evaluating Boolean formulas are impossible.
In this work, we identify parallelism as a key to distinguishing transformers from other architectures. While
recurrent architectures process their inputs serially, transformers allow independent interactions between the
input tokens, mediated by the inner products between query and key embeddings in self-attention units. We
leverage this property of self-attention to establish a formal connection between transformers and Massively
Parallel Computation (MPC) (Karloff et al., 2010). Concretely, we design transformers that simulate MPC
protocols (and vice versa), and in doing so, we exhibit a wide range of computational tasks that are solved by
logarithmic-depth transformers, including tasks that cannot be efficiently solved with other architectures
such as graph neural nets and recurrent models.
1.1
Our results
We advance the understanding of transformers’ representational capabilities with the following results.
1. The algorithmic capabilities and limitations of logarithmic-depth transformers are captured by the MPC
model (Section 3).
2. There is a simple sequential task that (i) is solved by (and, empirically, learned from data using) logarithmic-
depth transformers, but (ii) cannot be efficiently solved by several alternative architectures (Sections 4
and 5).
1
arXiv:2402.09268v1 [cs.LG] 14 Feb 2024
In more detail, our first collection of results, Theorems 3.1 and 3.4, show that any R-round MPC protocol
can be implemented by a transformer of depth O(R), and that any depth-L transformer can be simulated by an
O(L)-round MPC protocol. The former implies that several graph problems are solved by logarithmic-depth
transformers (Corollary 3.3); the latter implies the near-optimality of these transformers (Corollary 3.5)
conditional on a well-known conjecture about the limitations of MPC algorithms (Conjecture 2.4). A key
technical step (Lemma 3.2) shows how transformers can implement the simultaneous message-passing used in
MPC protocols to communicate between machines. While previous works (Sanford et al., 2023) have used
communication complexity to understand the representational limitations of self-attention layers, our results
show the benefits of the communication lens for understanding the strengths of transformers as well.
Our second set of results concern the k-hop induction heads task, a synthetic sequential task that draws
inspiration from the induction heads primitive of Elhage et al. (2021). The theoretical results of Section 4 prove
that depth L = Θ(log k) is necessary and sufficient for efficient transformer representation. An accompanying
empirical investigation reveals that transformers trained on the task obey the same threshold and recover
a similar model to the theoretical construction. In contrast, Section 5 illustrates that non-parallelizable
recurrent architectures—including state-space models like Mamba (Gu and Dao, 2023)—are unable to solve
the task in a size-efficient manner. Moreover, well-known transformer models with computationally-efficient
alternatives to self-attention, like Performer (Choromanski et al., 2022) and Longformer (Beltagy et al.,
2020), and shallow transformers with chain-of-thought prompting sacrifice their abilities to implement parallel
algorithms, as evidenced by their proven inability to solve this task.
1.2
Related work
Some of the types of lower bounds we sought in this work were inspired by the literature on depth-separation
for feed-forward neural networks (e.g., Eldan and Shamir, 2016; Daniely, 2017; Telgarsky, 2016), which exhibit
functions that are efficiently approximated by deep networks, but not by shallower networks.
Many theoretical approaches have been used to understand the representational capabilities of transformers
and self-attention units in various scaling regimes. Some works model (variants of) transformers as machines
for recognizing formal languages, such as the Dyck languages (Hahn, 2020; Bhattamishra et al., 2020; Yao
et al., 2021; Hao et al., 2022) and star-free regular languages (Angluin et al., 2023). These approaches reveal
inability of fixed-size transformers to handle arbitrarily long inputs. Other works show how transformers
can simulate finite-state automata Liu et al. (2022) with logarithmic depth, and Turing machines with
(unrolled) depth (or chain-of-thought length) scaling polynomially with total runtime Wei et al. (2021);
Malach (2023); Merrill and Sabharwal (2023b). However, it is unclear if these results are near optimal or
even transformer-specific.
Theoretical results about the limitations of constant-depth transformers have been articulated by way of
analogy to circuit complexity (Merrill and Sabharwal, 2023a; Merrill et al., 2022; Merrill and Sabharwal, 2022;
Strobl, 2023; Strobl et al., 2023), implying the inability of constant-depth transformers to solve tasks like
graph connectivity and Boolean formula evaluation. Other works characterize the representational capabilities
of one-layer transformers (Likhosherstov et al., 2021; Sanford et al., 2023), but these approaches do not apply
to deeper models. Sanford et al. study multi-headed attention using communication complexity, a framing
that informs this work’s connection to distributed computing.
The MPC model Karloff et al. (2010); Beame et al. (2017); Goodrich et al. (2011); Andoni et al. (2014);
Im et al. (2023) was introduced to study distributed computing frameworks such as MapReduce Dean and
Ghemawat (2004). A major goal is to design protocols that use few rounds of communication for setups
in which each machine’s local memory is sublinear in the input size. Many advances have been made in
MPC algorithms for important problems (see, e.g., Im et al., 2023, for a recent survey). However, a basic
problem that has resisted progress is connectivity in sparse graphs, where all MPC protocols in this memory
regime appear to require Ω(log n) rounds for input graphs on n vertices. Lower bounds in MPC and related
models were studied by Beame et al. (2017), Roughgarden et al. (2018), and Charikar et al. (2020). The
conjectured impossibility of o(log n)-round protocols for connectivity is now used as basis for conditional
lower bounds (Ghaffari et al., 2019).
Simulation of transformers by recurrent models (Oren et al., 2024) and simulation of graph neural nets
(GNNs) by transformers (Kim et al., 2022) offer some coarse-grain insight into the relationship between these
2
architectures, but separations are not implied by these previous works. Our connection between transformers
and MPC is most similar to that established by Loukas (2019) between GNNs and the Congest model of
distributed computation. Both works establish positive and negative results by identifying neural architectures
with communication protocols. In Section 5.1, we show that the MPC connection allows transformers solve
graph connectivity more efficiently than GNNs.
Our k-hop induction heads task is designed as a k-fold composition of its standard analogue (Elhage
et al., 2021). It is similar to a special case of the LEGO reasoning task Zhang et al. (2023), which reveals
the super-linear benefit of depth with respect to k; in our case, we theoretically and empirically exhibit an
exponential benefit. We also draw a connection to the well-studied problem of pointer-chasing (Papadimitriou
and Sipser, 1982; Duris et al., 1984; Nisan and Wigderson, 1993), which enables the proof of our separation
between parallel and serial architectures. Our fine-grained empirical interpretability analysis for synthetic
tasks draws inspiration from similar approaches for the analysis of sequential algorithms like sorting and
reversal (Li and McClelland, 2022).
2
Preliminaries
2.1
Transformers
We first define a self-attention head, the core primitive of a transformer. The softmax operator is softmax(v) =
(exp(v1), . . . , exp(vN))/ PN
j=1 exp(vj) for v ∈RN. We apply softmax to matrices A ∈RN×N row-wise, i.e.
softmax(A)i = softmax((Ai,1, . . . , Ai,N)).
Definition 2.1 (Self-attention head). A self-attention head is a mapping fQ,K,V : RN×m →RN×m defined
by
fQ,K,V (X) = softmax(Q(X)K(X)T)V (X)
and parameterized by row-wise query, key, and value embeddings Q, K, V : RN×m →RN×m (e.g., Q(X) =
(Q1(X1), . . . , QN(XN)). Let AttnN
m denote the set of all self-attention heads with embedding dimension m
and context length N.
A transformer composes L layers of H self-attention heads per layer, plus an output multi-layer perceptron
(MLP).
Definition 2.2 (Transformer). A transformer is a mapping T : RN×din →RN×dout specified by self-attention
heads (fℓ,h ∈AttnL
m)ℓ∈[L],h∈[H] and an element-wise output MLP ψ = (ψ1, . . . , ψN) : RN×m →RN×dout.
Upon input X ∈RN×din, the transformer computes intermediate embeddings X0, . . . , XL ∈RN×m with
X0 = X and
Xℓ= Xℓ−1 +
XH
h=1fℓ,h(Xℓ−1),
and returns T(X) = ψ(XL) as output. Let TransformerN
m,L,H,din,dout denote the set of all such transformers,
and TransformerN
m,L,H := TransformerN
m,L,H,1,1.
Modeling assumptions.
We treat the transformer as a computational model that permits arbitrary
element-wise computation, but restricts the manner in which multiple elements are processed together.
This manifests in our decisions to model query/key/value embeddings and MLPs as arbitrary functions
on the embedding space; Loukas (2019) employs a similar modeling assumption for GNNs. Note that the
element-wise embeddings and MLPs may be index-specific, obviating the need for positional embeddings.
Our theoretical results cover the scaling regime where the context length N is the main asymptotic
parameter; while the embedding dimension m, the number of heads H, and the depth L grow sub-linearly in
N. This reflects real-world trends in large-language models, where context length has sharply increased in
recent years.
Throughout, we assume all intermediate computations in transformers are represented by p-bit precision
numbers for p = Θ(log N). Limiting the precision is consistent with recent practice of using low-precision
arithmetic with transformers (e.g., Wang et al., 2022; Dettmers et al., 2022). We discuss this precision
assumption in greater detail in Appendix A.1, along with other minor technical assumptions (such as the
inclusion of a “start token” for mathematical convenience).
3
• Input = (Input1, . . . , Inputnin) ∈Znin
2p is distributed across local memories of machines 1 ≤i ≤⌈nin
s ⌉:
MachineIn(1)
i
= {(Inputι, ι) : ι ∈{(s−1)i+1, . . . , min {nin, si}}}.
• For round r = 1, . . . , R:
– Each machine i computes messages (MsgOut(r)
i,j )j=1,2,... to send to machines (Dest(r)
i,j )j=1,2,... as function of
MachineIn(r)
i
:
MachineOut(r)
i
= Localr,i(MachineIn(r)
i
) = {(MsgOut(r)
i,j , Dest(r)
i,j ) ∈Z
dj
2p × [q] : j = 1, 2, . . . };
X
jdj ≤s is ensured.
– All messages are simultaneously transmitted; the messages in local memory of machine i for round r + 1 are:
MachineIn(r+1)
i
= {(Msg, Src) : (Msg, i) ∈MachineOut(r)
Src};
X
(Msg,Src)∈MachineIn(r+1)
i
|Msg| ≤s is ensured.
• Output = f(Input) comes from
MachineIn(R+1)
i
= {(Outputι, Src) : ι ∈{(s −1)i+1, . . . , min {nout, si}}}
for 1 ≤i ≤⌈nout
s ⌉.
Figure 1: Formal execution of an MPC protocol for computing f : Znin
2p →Znout
2p
. (|Msg| is the number of
words in Msg.)
Masked transformers.
We also consider masked self-attention, where only certain inner products influence
the softmax output. Let Λ ∈{−∞, 0}N×N be a masking matrix with at least one zero entry in every row.
Then, a Λ-masked self-attention unit is defined by
f Λ
Q,K,V (X) = softmax(Q(X)K(X)T + Λ)V (X).
Let Λ-AttnN
m and Λ-TransformerN
m,L,H, respectively, denote the sets of all Λ-masked self-attention heads and
all transformers comprised of those heads. We define causally-masked transformers by MaskAttnN
m := Γ-AttnN
m
and MaskTransformerN
m,L,H := Γ-TransformerN
m,L,H, where Γ is the lower-triangular mask with Γi,j = 0 iff
i ≥j.
2.2
Massively Parallel Computation model
We use the definition of MPC from Andoni et al. (2018).
Definition 2.3 (MPC protocol). For any global and local memory constants γ, δ > 0, a (γ, δ)-MPC protocol
for a function f : Znin
2p →Znout
2p
specifies a distributed computing protocol for q = Θ(n1+γ−δ
in
) machines,
each with s = O(nδ
in) words1 of local memory to jointly compute f(Input) for any given Input ∈Znin
2p as
follows. The Input ∈Znin
2p is distributed across the local memories of the first ⌈nin/s⌉machines. Computation
proceeds in rounds. In each round, each machine computes an arbitrary function of its local memory to
prepare at most s words to send to other machines; messages are simultaneously transmitted, and the protocol
ensures that each machine receives at most s words at the end of the round. After the final round, the
Output = f(Input) ∈Znout
2p
is in the local memories of the first ⌈nout/s⌉machines. See Figure 1 for details.
Our negative results in Section 3.2 are conditional on the well-known “one-versus-two cycle” conjecture
(Beame et al., 2017; Roughgarden et al., 2018; Ghaffari et al., 2019).
1We assume the word size is p = Θ(log nin) bits. For convenience, we regard words as elements of Z2p (integers mod 2p).
4
Conjecture 2.4 (see, e.g., Ghaffari et al., 2019). For any γ > 0, δ < 1, and N, if π is an (γ, δ)-MPC
protocol that distinguishes a single cycle on N nodes and a union of two cycles each on N/2 nodes, then π
uses Ω(log N) rounds.
2.3
Graphs as sequential inputs
When providing a graph G = (V, E) as input to transformers or MPC protocols, we serialize G as a sequence
in [|V |]2|E| that encodes each edge as a pair of vertex tokens. The resulting transformer has N = 2|E| and
din = 1, and the resulting MPC protocol has nin = 2|E|.
3
Relating transformers and MPC
We coarsely characterize the computational power of transformers in a certain size regime by establishing a
bidirectional relationship between transformers and MPC. Theorems 3.1 and 3.4 show that any MPC protocol
can be simulated by a transformer, and vice versa. As corollaries (Corollaries 3.3 and 3.5), we obtain tight
upper and lower bounds on the depth of bounded-size transformers for computing connected components in
graphs.
3.1
Simulation of MPC protocols by transformers
The following theorem shows that any MPC protocol π with sublinear local memory can be simulated by
a transformer whose depth L is linear in the number of rounds R of π, and embedding dimension m is
polynomial in the local memory size s = O(N δ) of machines used by π.
Theorem 3.1. For constants 0 < γ < δ < 1 and any deterministic R-round (γ, δ)-MPC protocol π
on nin input words and nout ≤nin output words, there exists a transformer T ∈TransformerN
m,L,H with
N = nin, m = O(n4δ
in log nin), L = R + 1, H = O(log log nin) such that T(Input):nout = π(Input) for all
Input ∈ZN
2p.
The theorem provides a non-trivial construction in the strongly sub-linear local memory regime when
s = O(N 1/4−ϵ) for any ϵ > 0.2 Whether the simulation can be improved to m = O(N 1−ϵ′) for some ϵ′ > 0
whenever s = O(N 1−ϵ) is an interesting question for future work.
Theorem 3.1 proof overview.
At a high level, the proof in Appendix B.2 entails simulating each round
of parallel computation with a single-layer transformer and applying those constructions serially to Input.
The local computation on each machine (represented by MachineOut(r)
i
= Localr,i(MachineIn(r)
i )) is directly
encoded using element-wise query/key/value embeddings.
The crux of the proof involves the simulation of a routing protocol to determine MachineIn(r+1) from
MachineOut(r). We construct a self-attention unit that ensures that an encoding of a sequence of addressed
messages from each machine are properly routed to their destinations.3
For any message size β, message count bound s, and number of tokens N, we say that (Sent, Rcvd) ∈
RN×m × RN×m is a valid (β, s)-routing if, for each i ∈[N], the i-th row of Sent (resp. Rcvd) is the vector
encoding of some Senti ⊂Zβ
2p × [N] (resp. Rcvdi ⊂Zβ
2p × [N]) such that
Rcvdi = {(Msg, Src) : (Msg, i) ∈SentSrc} ,
and each of Rcvdi and Senti has cardinality at most s.4
Lemma 3.2. For any β, s, N ∈N, there exists a transformer routeβ,s ∈TransformerN
m,1,1 with m =
O(s4β log N) satisfying routeβ,s(Sent) = Rcvd for any valid (β, s)-routing (Sent, Rcvd).
2Applying Theorem 3.1 when δ ≥1
4 yields transformers with embedding dimension m ≥N, which trivializes the transformer
architecture and negates any advantages of depth under our MLP universality assumption. This is due to the fact a transformer
with N-dimensional embeddings could aggregate the entire input sequence X ∈RN in a single embedding and use its output
MLP to compute any arbitrary function on that input.
3This routing between machines uses the all-pairs structure of self-attention and may not admit a subquadratic approximation.
4We abuse notation by writing Dest ∈Senti to mean there exists some Msg such that (Msg, Dest) ∈Senti.
5
The proof of Lemma 3.2 appears in Appendix B.1 and combines two key techniques: sparse propagation
and multiple hashing. The former is a simple variant of the “sparse averaging” task of Sanford et al. (2023),
which simultaneously computes N averages over subsets of inputs; this task is solved a single self-attention
head with small embedding dimension (Proposition B.1). Using sparse propagation, we construct a self-
attention head that averages the ≤s encodings of each RcvdSrc for every Src ∈Rcvdi. In order to ensure that
we can decode that average of encodings, we apply error-correction by encoding each Outputi in a sparse and
redundant manner, where each outgoing messages appears as multiple copies of the same addressed “packet.”
Application: connectivity with log-depth transformers.
As an immediate consequence of Theorem 3.1,
any graph problem solvable with a logarithmic number of rounds of MPC computation (and local memory s)
is also computable by a logarithmic depth transformer (and embedding dimension ˜O(s4)). The following
result—which bounds transformer depth needed to compute connected components of a graph G—follows
from Theorem 6.2 of Coy and Czumaj (2022), which derandomizes an MPC algorithm of Behnezhad et al.
(2019), and Theorem 3.1.
Corollary 3.3. For any constant ϵ ∈(0, 1) and any D ≤N, there exists a transformer in TransformerN
m,L,H
with m = O(N ϵ), H = O(log log N), and L = O(log D) that identifies the connected components of any input
graph G = (V, E) with |V |, |E| = O(N) where each connected component has diameter at most D.
Coy and Czumaj also give efficient MPC algorithms for other related problems (e.g., spanning forest), so
we obtain efficient transformers for these problems, too (Appendix B.3).
3.2
Simulation of transformers by MPC protocols
The following theorem shows that MPC protocols can simulate transformers and prove depth lower bounds
on transformers, conditioned on Conjecture 2.4. We get, as a corollary, the conditional optimality of the
transformer depth bound in Corollary 3.3.
Theorem 3.4. For any transformer T ∈TransformerN
m,L,H (or Λ-TransformerN
m,L,H) with mH = O(N δ) for
δ ∈(0, 1) and any δ′ ∈(δ, 1), there exists a O(
L
δ′−δ)-round (1+δ′, δ′)-MPC protocol with q = O(N 2) machines
with s = O(N δ′) local memory for computing T.
Theorem 3.4 demonstrates that the algorithmic capabilities of transformers are no stronger than those of
MPC protocols with a quadratic scaling in the number of machines. While Theorems 3.1 and 3.4 do not
jointly provide a sharp characterization of the two computational models, the reductions are tight enough to
provide strong evidence for the optimality of the connected components construction of Corollary 3.3.
Theorem 3.4 proof overview.
At a high-level, the proof constructs an MPC protocol that simulates
a self-attention layer by separating the computation of MLPs and attention matrices into three separate
categories of machines.
• Each input token is provided to its own token machine, responsible for preparing the query/key/value
embeddings.
• Each pair of tokens is associated with an inner product machine that will compute the inner product
between their respective query and key embeddings.
• Propagation machines ensure that embeddings are routed to the proper inner product machine and compute
outputs of each softmax unit.
The proof gives the communication protocol for these machines, shows how they simulate a layer of self-
attention in O(1/(δ′ −δ)) rounds, and establishes the sufficiency of O(N 2) machines with O(N δ′) local
memory.
Application: conditional optimality of Corollary 3.3.
Assuming the well-established Conjecture 2.4,
we prove an Ω(log D) lower bound on the depth of parameter-efficient transformers for determining connectivity
of graphs where connected components may have diameter up to D.
Corollary 3.5. Let ϵ ∈(0, 1) be any constant, and let D ≥N ϵ. Assume Conjecture 2.4, and suppose there
exists T ∈TransformerN
m,L,H with mH = O(D1−ϵ) that decides connectivity of any input graph with connected
components having diameter ≤D. Then L = Ω(log D).
6
4
Transformers for k-hop induction heads
We complement the generality of Section 3 by studying, both empirically and theoretically, a specific toy
sequential modeling task which will also serve (in Section 5) as a problem to separate the representational
capabilities of transformers from that of other neural architectures.
This task, called the k-hop induction heads task, draws inspiration from the original induction heads task
defined and analyzed on trained language models and in synthetic environments by Elhage et al. (2021) (see
also Bietti et al., 2023). The standard induction heads task completes bigrams auto-regressively by predicting
the token that follows the last previous occurrence of the final token in the sequence. For example, given the
input X = baebcabebdea, the standard induction heads task is to complete the final bigram by predicting b
for the final token.
The k-hop induction heads tasks generalizes this mechanism by repeatedly using the completion of a
bigram to determine the next bigram to complete. In the previous example, the 2-hop induction heads task
is to predict c for the final token:
baebcabebdea.
Definition 4.1. For any finite alphabet Σ, define the map hopk : ΣN →(Σ ∪{⊥})N by hopk(X)i = Xfindk
X(i)
if findk
X(i) ̸= 0 and ⊥otherwise, where
find1
X(i) = max({0} ∪{j ∈N : j ≤i, Xj−1 = Xi});
findk
X(i) = find1
X(findk−1
X
(i))
for k ≥2.
The k-hop induction heads task is to compute, for each i = 1, . . . , N, the value of hopk(X)i from (X1, . . . , Xi).
We note a similarity to the LEGO tasks of Zhang et al. (2023), who empirically study the ability of
transformers to learn sequential operations on Abelian groups and observe the ability to perform more
operations than the depth of the network.
4.1
Log-depth transformer for k-hop induction heads
Although hopk appears to requires k steps to solve, we show that it is solved by a transformer of depth
O(log k).
Theorem 4.2. For any k ∈N and alphabet Σ with |Σ| ≤N, there exists T ∈MaskTransformerN
m,L,H that
computes hopk : ΣN →(Σ ∪{⊥})N with m = O(1), L = ⌊log2 k⌋+ 2, and H = 1.
In contrast to Corollary 3.3, this construction has constant embedding dimension and is achieved by a
causally-masked transformer. As such, its proof in Appendix D.1 depends on other techniques that exploit
the simplicity of the problem and build on the induction heads construction of Bietti et al. (2023), rather
than simply applying Theorem 3.1.
We give evidence for the optimality of this construction by proving a conditional lower bound using
Theorem 3.4, as was done in Corollary 3.5.
Corollary 4.3. Assuming Conjecture 2.4, for any constants ξ ∈(0, 1/2] and ϵ ∈(0, 1), and any even
k = Θ(N ξ), every transformer T ∈MaskTransformerN
m,L,H with mH = O(k1−ϵ) that computes hopk has depth
L = Ω(log k).
4.2
Log-depth transformer learned from data
We empirically assess whether the representational trade-offs elucidated by tasks efficiently solved by
parallelizable algorithms have implications for optimization and generalization properties of transformers. To
that end, we trained auto-regressive transformer architectures of varying sizes to solve hopk(X) for a variety
of values of k in order to understand how changing depth impacted the performance of the learned models,
the goal being to verify the sufficiency of logarithmic depth, just as in our theory.
7
1
2
4
8
16
k
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Error
L-layer transformer token-wise classification error on hopk
L = 2
L = 3
L = 4
L = 5
L = 6
Figure 2: Evaluation of transformers of depths L ∈{2, 3, 4, 5, 6} trained on a mixture of hopk for k ∈
{0, . . . , 16} evaluated on n = 100 samples of size N = 100 from each hopk. Incrementing depth approximately
doubles the largest k such that hopk is learnable with small error.
In brief, we trained transformers with 500K to 5M parameters and depths {2, 3, 4, 5, 6} with Adam
to solve hopk(X) for k ∈{0, . . . , 16} with context length |N| = 100 and alphabet size |Σ| = 4.
We
trained the transformers in a multi-task setting, where a single model was trained to predict the sequence
hopk(X) auto-regressively when provided with X and k drawn at random. Further experimental details
can be found in Appendix G.1, and the experimental code is available at https://github.com/chsanford/
hop-induction-heads.
We found that transformers are indeed capable of learning hopk given sufficient training time, and that the
largest learnable k grows exponentially with the depth. As can be seen in Figure 2, a six-layer neural network
performs well on all k ≤16, a five-layer on k ≤8, a four-layer on k ≤4, and so forth. We further explore
these experimental results in Appendix G.2 and observe a performance threshold appears to specifically lie at
⌊log2 k⌋+ 2 that coincides with Theorem 4.2. This logarithmic dependence of the depth on k persists in a
larger-width regime, which is explored in Appendix G.3. In the finite sample regime where neural networks
are prone to overfit, our investigations in Appendix G.5 note improved generalization in deeper models, which
suggests that deeper models have a favorable inductive bias for tasks like hopk.
Moreover, the learned models are surprisingly interpretable. We examined the activation patterns of
attention matrices, and found close correspondences to useful intermediate products such as findj
X. Taken
together, these indicate that the learned models mechanistically resemble the construction employed in the
proof of Theorem 4.2. See Appendix G.4 for our investigation of model interpretability.
5
Separations between transformers and alternative architectures
Sections 3 and 4 characterize the representational capability of transformers by providing algorithmic problems
they can solve with logarithmic depth and small polynomial or constant width. In contrast, other well-known
architectures are unable to solve those same problems in a parameter-efficient manner. This section provides
lower bounds on the parameter complexity of graph neural networks (GNNs), recurrent neural architectures,
transformers with computationally efficient alternatives to softmax self-attention, and single-layer transformers
with autoregressive chain-of-thought tokens needed to solve graph connectivity and the k-hop task.
8
5.1
GNNs need polynomial depth for graph connectivity
The bidirectional relationship between transformers and MPC draws inspiration from past work drawing a
similar connection between message passing graph neural networks (GNNmp) and the Congest distributed
computing model Loukas (2019). Their computation model of GNNmp for width m and depth L closely
resembles our TransformerN
m,L,H in providing a general framework for the analysis of graph neural networks
by allowing unbounded computation in each vertex with bounded communication on edges. On some input
graph G, vertices send neighbors messages of size at most m—which are aggregated and crafted into new
messages with MLPs—over L rounds of communication.
By restating Corollary 4.2 of Loukas (2019), we demonstrate a sharp contrast in the abilities of GNNs
and transformers to solve graph algorithmic tasks.
Theorem 5.1 (Corollary 4.2 of Loukas (2019)). There exists a graph G with N edges such that any GNNmp
with width m and depth L that determines whether an input subgraph H either (1) is connected or (2) forms
a spanning tree of G requires L√m = ˜Ω(N 1/4).
While Corollaries 3.3 and B.8 demonstrate the ability of transformers to determine whether any input
graph is connected5 or to identify a spanning tree with logarithmic depth and small polynomial width (i.e.
m = O(N ϵ)), GNNs require depth L = ˜Ω(N 1/4−ϵ/2) in the same regime. This gap is explainable by the fact
that transformers on graph inputs G are not bound to pass messages exclusively along the edges of G. By
“rewiring” the graphical structure in each layer, transformers can perform aggregation and “pointer passing”
tasks with greater parametric ease than GNNs.
5.2
Suboptimality of recurrent architectures for hopk
The logarithmic-depth and constant-width transformer implementation of hopk in Theorem 4.2 cannot be
replicated by recurrent neural architectures (Chung et al., 2014; Bengio et al., 1994; Turkoglu et al., 2021),
including not just multi-layer recurrent neural networks (RNNs) but any sequential prediction procedure
equivalent to them at inference time, which includes state space models such as Mamba (Gu and Dao, 2023).
We first consider a family of multi-layer RNNs of depth L and width m, consisting of arbitrary MLP units
gℓ: Rm×m →Rm×m, which on input X ∈RN×din produce output Y ∈RN×dout as follows using intermediates
X = Z0, Z1, . . . , ZL−1, ZL = Y ∈RN×m6and hidden states H1, . . . , HL ∈{0, 1}N×m with Hℓ
0 = ⃗0:
(Zℓ
i , Hℓ
i ) = gℓ(Zℓ−1
i
, Hℓ
i−1), ∀i ∈[N], ℓ∈[L].
We provide a polynomial bound on the width and depth of a multi-layer RNN solving hopk.
Corollary 5.2. A multi-layer RNN of depth L and width m as above with YN = hopk(X)N satisfies either
L ≥k or m = Ω( N
k6 ).
In contrast to Theorem 4.2, which demonstrates that depth O(log k) transformers with constant width
suffice to solve hopk for any k, Corollary 5.2 demonstrates that all multi-layer RNNs with width O(N 1/7)
require depth k when k = O(N 1/7).
Mamba (Gu and Dao, 2023) can be seen as the combination of three ideas: (1) a continuous-time
dynamics model of sequential prediction, powerful enough to model Kalman filters, hidden markov models,
and many others; (2) a family of time-discretization schemes; (3) an unrolling technique to enable efficient
linear-time training, using ideas similar to FlashAttention (Dao et al., 2022). Ultimately, at inference time,
the time-discretization step results in an RNN (see Gu and Dao, 2023, Algorithm 2 and Theorem 1), and is
therefore directly handled by Corollary 5.2.
This corollary is a near immediate application of a communication complexity fact about the hardness
of solving multi-player pointer-chasing problems with limited communication among players (Guha and
McGregor, 2009; Assadi and N, 2021). We provide the communication model and this result in Appendix E.1,
and the reductions necessary to prove the above hardness results in Appendix E.2.
5While the problem of subgraph connectivity for GNNs may at first glance appear more difficult than general graph connectivity
for transformers, an implementation of this exact task can be implemented by modifying the protocol Corollary 3.3 to remove all
edges from the graph that do not belong to H.
6We assume that din, dout ≤m and treat X and Y as if they are padded with zeros.
9
5.3
Suboptimality of sub-quadratic attention transformers for hopk
Due to the quadratic computational cost of computing the attention matrix softmax(Q(X)K(X)T ) ∈RN×N
and the continued desire for ever-larger context lengths, there is substantial interest in improving the
computational complexity of the transformer architecture while preserving its expressive capabilities and
inductive biases. As a result, a rich literature has emerged that proposes computationally-efficient alternatives
to standard softmax attention. In this section, we demonstrate how several representative examples of
sub-quadratic attention mechanisms lose the ability to perform efficient parallel computation under a
logarithmic-depth scaling.
Kernel-based sub-quadratic attention.
One approach to computationally-efficient approximation of
transformers are kernel-based sub-quadratic attention mechanisms such as Performer (Choromanski et al., 2022),
and Poly-Sketchformer (Kacham et al., 2023). Both approximate the attention matrix softmax(Q(X)K(X)T)
with a low-rank matrix Q′(X)K′(X)T where Q′, K′ : Rm →Rm′ are applied element-wise. For sufficiently
small m′ ≪N, Q′(X)K′(X)TV (X) can be computed efficiently by first computing K′(X)TV (X) ∈Rm′×m,
bounding the total runtime as O(Nmm′), rather than O(N 2m).
Let KernelFormerN
m,m′,L,H denote all H-headed L-layer transformer whose softmax attention modules are
replaced by kernel-based sub-quadratic attention. We demonstrate the limitations of KernelFormerN
m,m′,L,H
by showing that, unlike TransformerN
m,L,H, they have no depth-efficient implementation of hopk.
Corollary 5.3. Any T ∈KernelFormerN
m,m′,L,H with T(X)N = hopk(X)N satisfies either L ≥k or mm′Hp =
Ω( N
k6 ).
Under a parameter-efficient regime where mpHL = O(N ϵ), solving hopk for k = Θ(N ϵ) necessitates
kernel feature dimension m′ = Ω(N 1−9ϵ), which forces each attention unit to compute an N × N 1−9ϵ matrix,
yielding a nearly quadratic runtime. We prove Corollary 5.3 in Appendix E.3 using a similar pointer chasing
reduction.
Masking-based sub-quadratic attention.
Another method that reduces the computational cost of
transformers is to used masked models of Λ-TransformerN
m,L,H for a sparse mask Λ.
The Longformer
architecture (Beltagy et al., 2020) introduces a particular masked architecture that combines sliding windows
with sparse unmasked global tokens. Put concretely, for window radius w and global frequency g, let
Λw,g ∈{−∞, 0}N×N be masking matrix with
Λw,g
i,j =
(
0
if |i −j| ≤w or j ≡0
(mod g),
−∞
otherwise.
Then, the output of a single unit of Λw,g-masked attention is computable in time O((w + N
g )Nm).
Corollary 5.4. Any T ∈Λw,g-AttnN
m,L,H with T(X)N = hopk(X)N satisfies either L ≥k or (w+ N
gk)mHp =
Ω( N
k6 ).
Like kernel-based attention, sparsely-masked attention models fail to efficiently compute hopk. Similarly,
in the same parameter-efficient regime, a Longformer must have either w = Ω(N 1−9ϵ) or g = O(N 9ϵ), which
jointly ensures that the masked matrix has at least Ω(N 2−9ϵ) entries and diminishes any computational
advantages. This proof also appears in Appendix E.3.
5.4
Limitations of 1-layer transformers with chain-of-thought
While most of the paper considers transformers as sequence-to-sequence models, we can also frame them as
auto-regressive models performing next-token-prediction with chain-of-thought prompting. In this regime, a
single causally-masked transformer aims to compute a function of its input by repeatedly predicting the next
token, appending previously predicted tokens to the end of the input. In doing so, a function is computable if
there exists an intermediate chain-of-thought produced by the model that eventually reaches the answer.
10
Definition 5.5. We say that T ∈MaskTransformerN+NCoT
m,L,H
computes f : ΣN+NCoT →ΣN, where the
additional N tokens denote chain-of-thought, if for every X ∈dom(f), there exists XCoT ∈ΣNCoT such that
T(X ◦XCoT)N:N+NCoT = (XCoT ◦f(X)).
The theoretical capabilities of chain-of-thought augmented transformers to simulate finite-state automata
and Turing machines have been studied (Malach, 2023; Merrill and Sabharwal, 2023b), but the comparative
capabilities of shallow models with chain-of-thought prompting and deep sequential models are unknown. In
contrast to the fact that any transformer with NCoT tokens can be simulated by a sequential model with
depth scaled by NCoT, we show that deep transformers cannot necessarily be efficiently simulated by shallow
chain-of-thought models. We do so by demonstrating that a linear amount of chain-of-thought prompting in
k is necessary to solve hopk(X)N, and also sufficient.
Corollary 5.6. Any T ∈MaskTransformerN+NCoT
m,1,H
that computes hopk(X)N with NCoT tokens of chain-of-
thought requires either NCoT ≥k or mHp = Ω( N
k6 ).
The proof appears in Appendix E.4. For future work, it remains to consider the comparative powers of
chain-of-thought models of depths greater than one.
6
Conclusion and future work
This work highlights parallelism as a central feature of transformers that sets them apart from other neural
architectures. The focus on the log-depth and sublinear-width regime and specific computational tasks allows
us to accentuate the benefits of parallelism, even for tasks like k-hop that appear inherently serial at first
glance.
There is some efficiency loss in the “compilation” of MPC protocols to transformers that we hope to
understand better in future work. Furthermore, although we have empirically demonstrated the learnability
of transformers that exploit parallelism in crucial ways, a theoretical understanding of learning such solutions
remains an open question.
References
Alekh Agarwal, Olivier Chapelle, Miroslav Dudík, and John Langford. A reliable effective terascale linear
learning system. Journal of Machine Learning Research, 15(1):1111–1133, 2014.
Alexandr Andoni, Aleksandar Nikolov, Krzysztof Onak, and Grigory Yaroslavtsev. Parallel algorithms
for geometric graph problems. In Proceedings of the forty-sixth annual ACM symposium on Theory of
computing, pages 574–583, 2014.
Alexandr Andoni, Zhao Song, Clifford Stein, Zhengyu Wang, and Peilin Zhong. Parallel graph connectivity in
log diameter rounds. In 2018 IEEE 59th Annual Symposium on Foundations of Computer Science (FOCS).
IEEE, October 2018. doi: 10.1109/focs.2018.00070. URL http://dx.doi.org/10.1109/FOCS.2018.00070.
Dana Angluin, David Chiang, and Andy Yang. Masked hard-attention transformers and boolean rasp
recognize exactly the star-free languages, 2023.
Sepehr Assadi and Vishvajeet N. Graph streaming lower bounds for parameter estimation and property
testing via a streaming xor lemma.
In Proceedings of the 53rd Annual ACM SIGACT Symposium
on Theory of Computing, STOC ’21. ACM, June 2021.
doi: 10.1145/3406325.3451110.
URL http:
//dx.doi.org/10.1145/3406325.3451110.
Paul Beame, Paraschos Koutris, and Dan Suciu. Communication steps for parallel query processing. Journal
of the ACM (JACM), 64(6):1–58, 2017.
Soheil Behnezhad, Sebastian Brandt, Mahsa Derakhshan, Manuela Fischer, MohammadTaghi Hajiaghayi,
Richard M Karp, and Jara Uitto. Massively parallel computation of matching and mis in sparse graphs. In
Proceedings of the 2019 ACM Symposium on Principles of Distributed Computing, pages 481–490, 2019.
11
Iz Beltagy, Matthew E. Peters, and Arman Cohan. Longformer: The long-document transformer, 2020.
Y. Bengio, P. Simard, and P. Frasconi. Learning long-term dependencies with gradient descent is difficult.
IEEE Transactions on Neural Networks, 5(2):157–166, 1994. doi: 10.1109/72.279181.
Satwik Bhattamishra, Kabir Ahuja, and Navin Goyal. On the ability and limitations of transformers to
recognize formal languages. In Proceedings of the 2020 Conference on Empirical Methods in Natural
Language Processing, 2020.
Alberto Bietti, Vivien Cabannes, Diane Bouchacourt, Herve Jegou, and Leon Bottou. Birth of a transformer:
A memory viewpoint, 2023.
Moses Charikar, Weiyun Ma, and Li-Yang Tan. New lower bounds for massively parallel computation from
query complexity, 2020.
Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos,
Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, David Belanger, Lucy Colwell, and Adrian
Weller. Rethinking attention with performers, 2022.
Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated
recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014.
Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D Manning. What does bert look at? an
analysis of bert’s attention. arXiv preprint arXiv:1906.04341, 2019.
Sam Coy and Artur Czumaj. Deterministic massively parallel connectivity. In Proceedings of the 54th Annual
ACM SIGACT Symposium on Theory of Computing, STOC 2022, page 162–175, New York, NY, USA,
2022. Association for Computing Machinery. ISBN 9781450392648. doi: 10.1145/3519935.3520055. URL
https://doi.org/10.1145/3519935.3520055.
Amit Daniely. Depth separation for neural networks. In Satyen Kale and Ohad Shamir, editors, Proceedings
of the 2017 Conference on Learning Theory, volume 65 of Proceedings of Machine Learning Research, pages
690–696. PMLR, 07–10 Jul 2017. URL https://proceedings.mlr.press/v65/daniely17a.html.
Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-
efficient exact attention with io-awareness. In NeurIPS, 2022.
Jeffrey Dean and Sanjay Ghemawat. Mapreduce: Simplified data processing on large clusters. In OSDI,
pages 137–150, 2004.
Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. Llm.int8(): 8-bit matrix multiplication
for transformers at scale. In Advances in Neural Information Processing Systems, volume 35, 2022.
Pavol Duris, Zvi Galil, and Georg Schnitger. Lower bounds on communication complexity. In Proceedings of
the Sixteenth Annual ACM Symposium on Theory of Computing, page 81–91, 1984.
Ronen Eldan and Ohad Shamir. The power of depth for feedforward neural networks. In Vitaly Feldman,
Alexander Rakhlin, and Ohad Shamir, editors, 29th Annual Conference on Learning Theory, volume 49 of
Proceedings of Machine Learning Research, pages 907–940, Columbia University, New York, New York,
USA, 23–26 Jun 2016. PMLR. URL https://proceedings.mlr.press/v49/eldan16.html.
Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell,
Yuntao Bai, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds,
Danny Hernandez, Andy Jones, Jackson Kernion, Liane Lovitt, Kamal Ndousse, Dario Amodei, Tom Brown,
Jack Clark, Jared Kaplan, Sam McCandlish, and Chris Olah. A mathematical framework for transformer
circuits. Transformer Circuits Thread, 2021. https://transformer-circuits.pub/2021/framework/index.html.
12
Mohsen Ghaffari, Fabian Kuhn, and Jara Uitto. Conditional hardness results for massively parallel computation
from distributed lower bounds. In IEEE 60th Annual Symposium on Foundations of Computer Science,
pages 1650–1663, 11 2019. doi: 10.1109/FOCS.2019.00097.
Michael T Goodrich, Nodari Sitchinava, and Qin Zhang. Sorting, searching, and simulation in the mapreduce
framework. In International Symposium on Algorithms and Computation, pages 374–383. Springer, 2011.
Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces, 2023.
Sudipto Guha and Andrew McGregor. Stream order and order statistics: Quantile estimation in random-
order streams. SIAM Journal on Computing, 38(5):2044–2059, 2009. doi: 10.1137/07069328X. URL
https://doi.org/10.1137/07069328X.
Michael Hahn. Theoretical limitations of self-attention in neural sequence models. Trans. Assoc. Comput.
Linguistics, 8:156–171, 2020. doi: 10.1162/tacl\_{a}{\_{0}{0}{3}}{0}6. URL https://doi.org/10.
1162/tacl_a_00306.
Yiding Hao, Dana Angluin, and Robert Frank. Formal language recognition by hard attention transformers:
Perspectives from circuit complexity. Trans. Assoc. Comput. Linguistics, 10:800–810, 2022. URL https:
//transacl.org/ojs/index.php/tacl/article/view/3765.
Sungjin Im, Ravi Kumar, Silvio Lattanzi, Benjamin Moseley, Sergei Vassilvitskii, et al. Massively parallel
computation: Algorithms and applications. Foundations and Trends® in Optimization, 5(4):340–417, 2023.
John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn
Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure
prediction with alphafold. Nature, 596(7873):583–589, 2021.
Praneeth Kacham, Vahab Mirrokni, and Peilin Zhong. Polysketchformer: Fast transformers via sketches for
polynomial kernels, 2023.
Howard Karloff, Siddharth Suri, and Sergei Vassilvitskii. A model of computation for mapreduce. In
Twenty-first Annual ACM-SIAM Symposium on Discrete Algorithms, pages 938–948, 12 2010. doi: 10.1137/
1.9781611973075.76.
Jinwoo Kim, Tien Dat Nguyen, Seonwoo Min, Sungjun Cho, Moontae Lee, Honglak Lee, and Seunghoon
Hong. Pure transformers are powerful graph learners, 2022.
Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2014.
Yuxuan Li and James L. McClelland. Systematic generalization and emergent structures in transformers
trained on structured tasks, 2022.
Valerii Likhosherstov, Krzysztof Choromanski, and Adrian Weller. On the expressive power of self-attention
matrices. arXiv preprint arXiv:2106.03764, 2021.
Bingbin Liu, Jordan T. Ash, Surbhi Goel, Akshay Krishnamurthy, and Cyril Zhang. Transformers learn
shortcuts to automata, 2022.
Andreas Loukas. What graph neural networks cannot learn: depth vs width. arXiv preprint arXiv:1907.03199,
2019.
Eran Malach. Auto-regressive next-token predictors are universal learners, 2023.
William Merrill and Ashish Sabharwal. A logic for expressing log-precision transformers, 2022.
William Merrill and Ashish Sabharwal. The parallelism tradeoff: Limitations of log-precision transformers.
Transactions of the Association for Computational Linguistics, 11:531–545, 2023a. ISSN 2307-387X. doi:
10.1162/tacl_a_00562. URL http://dx.doi.org/10.1162/tacl_a_00562.
13
William Merrill and Ashish Sabharwal. The expressive power of transformers with chain of thought, 2023b.
William Merrill, Ashish Sabharwal, and Noah A. Smith. Saturated transformers are constant-depth threshold
circuits. Transactions of the Association for Computational Linguistics, 10:843–856, 2022. ISSN 2307-387X.
doi: 10.1162/tacl_a_00493. URL http://dx.doi.org/10.1162/tacl_a_00493.
MPICH. Mpi allreduce, 2023. URL https://www.mpich.org/static/docs/latest/www3/MPI_Allreduce.
html.
Noam Nisan and Avi Wigderson. Rounds in communication complexity revisited. SIAM Journal on Computing,
22(1):211–219, 1993. doi: 10.1137/0222016. URL https://doi.org/10.1137/0222016.
Matanel Oren, Michael Hassid, Yossi Adi, and Roy Schwartz. Transformers are multi-state rnns, 2024.
Christos H. Papadimitriou and Michael Sipser. Communication complexity. In Proceedings of the Fourteenth
Annual ACM Symposium on Theory of Computing, page 196–200, 1982.
Jorge Pérez, Pablo Barceló, and Javier Marinkovic. Attention is turing complete. Journal of Machine
Learning Research, 22(1):3463–3497, 2021.
Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models
are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.
Anna Rogers, Olga Kovaleva, and Anna Rumshisky. A primer in bertology: What we know about how bert
works. Transactions of the Association for Computational Linguistics, 8:842–866, 2021.
Tim Roughgarden, Sergei Vassilvitskii, and Joshua Wang. Shuffles and circuits (on lower bounds for modern
parallel computation). Journal of the ACM, 65:1–24, 11 2018. doi: 10.1145/3232536.
Clayton Sanford, Daniel Hsu, and Matus Telgarsky. Representational strengths and limitations of transformers,
2023.
Lena Strobl. Average-hard attention transformers are constant-depth uniform threshold circuits, 2023.
Lena Strobl, William Merrill, Gail Weiss, David Chiang, and Dana Angluin. Transformers as recognizers of
formal languages: A survey on expressivity, 2023.
Matus Telgarsky. Benefits of depth in neural networks. In Vitaly Feldman, Alexander Rakhlin, and Ohad
Shamir, editors, 29th Annual Conference on Learning Theory, volume 49 of Proceedings of Machine Learning
Research, pages 1517–1539, Columbia University, New York, New York, USA, 23–26 Jun 2016. PMLR.
URL https://proceedings.mlr.press/v49/telgarsky16.html.
Mehmet Ozgur Turkoglu, Stefano D’Aronco, Jan Dirk Wegner, and Konrad Schindler. Gating revisited: Deep
multi-layer rnns that can be trained. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44
(8):4081–4092, 2021.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser,
and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems 30,
2017.
Ziwei Wang, Changyuan Wang, Xiuwei Xu, Jie Zhou, and Jiwen Lu. Quantformer: Learning extremely
low-precision vision transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022.
Colin Wei, Yining Chen, and Tengyu Ma. Statistically meaningful approximation: a case study on approxi-
mating turing machines with transformers, 2021.
14
Shunyu Yao, Binghui Peng, Christos H. Papadimitriou, and Karthik Narasimhan. Self-attention networks can
process bounded hierarchical languages. In Proceedings of the 59th Annual Meeting of the Association for
Computational Linguistics and the 11th International Joint Conference on Natural Language Processing,
2021.
Chulhee Yun, Srinadh Bhojanapalli, Ankit Singh Rawat, Sashank Reddi, and Sanjiv Kumar. Are transformers
universal approximators of sequence-to-sequence functions?
In International Conference on Learning
Representations, 2020.
Yi Zhang, Arturs Backurs, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, and Tal Wagner. Unveiling
transformers with lego: a synthetic reasoning task, 2023.
15
A
Supplemental Preliminaries
A.1
Further details about transformers
We discuss a few minor technicalities and modifications of the self-attention unit (Definition 2.1) and
transformer model (Definition 2.2) defined in Section 2.1 that are necessary for readers looking for a
comprehensive understanding of the proofs of our theoretical results.
Fixed-bit precision arithmetic.
As discussed in Section 2.1, we assume that all numbers that appear in
the intermediate products and outputs of self-attentions are representable with p-bit precision arithmetic,
where p = Θ(log N). While the details of fixed-precision arithmetic will be uninteresting to most readers, it
is necessary to explain precisely what we mean in order to ensure that proofs of results like Theorem 3.4 are
sound. Throughout the paper, we allow p to depend on of constants, such as γ, δ, and ϵ.
Concretely, we assume that all query, key, and value embeddings Q(X), K(X), V (X) evaluated on all
inputs contain scalar values z ∈R that are polynomially bounded (i.e. |z| ≤exp(O(p)) = N ζ for sufficiently
large constant exponent ζ > 0) and are inverse-polynomially discretized (i.e. z · N ζ ∈Z). Depending on the
desired exponent ζ, some p = Θ(log N) can be chosen to guarantee this property. While we do not formally
analyze the precision needed to approximate the particular embeddings employed by our proofs, we note
that our recurring sinusoidal embeddings (e.g. Lemma D.1) can be discretized without losing their central
properties and that discretizations of the restricted isometry embeddings of Proposition B.1 are analyzed by
Sanford et al. (2023).
Rather than stipulating a particular bounded-precision implementation that computes the output of a self-
attention unit must be implemented, we specify a rounding constraint that any computational implementation
of a self-attention unit must satisfy. Precisely, we require that any output round to the same inverse-polynomial
discretization as the true mathematical attention.
Definition A.1. For a self-attention unit f ∈AttnN
m, let ˆf be an finite-precision implementation of that unit.
We say that ˆf is a valid implementation if
sup
X∈RN×m
f(X) −ˆf(X)
∞= O
1
2p
.
This definition is only to establishing the fact that self-attention units with sufficient margins can precisely
compute hardmax outputs in Lemma A.2 and to showing that MPC models can indeed compute the outputs
precisely in Theorem 3.4.
Hardmax attention.
While we exclusively consider attention units with the softmax, our constructions
periodically rely on the exact computation of averages of embeddings. We define the hardmax operator to
allow the consideration of discrete averaging operations. For some v ∈RN, let
hardmax(X)i =
(
1
|Imax(v)|,
if i ∈Imax(v)
0
otherwise,
where Imax(v) = {i ∈[N] : vi = maxi′ vi′}.
We show that bounded-precision softmax self-attention units that satisfy a margin property can be
modified slightly to have identical outputs to an analogous hardmax unit.
Lemma A.2. Let f ∈AttnN
m be a self-attention unit with precision p = Θ(log N) and embedding functions
Q, K, V such that for some fixed 1 ≥ξ = N −O(1) and every X ∈RN×m and i ∈[N]:
A(X)i,i′ ≤max
i′′ A(X)i,i′′ −ξ, ∀i′ ̸∈Imax(A(X)i),
where A(X) = Q(X)K(X)T. Then there exists a self-attention unit f ′ ∈AttnN
m with a valid p′-bit implemen-
tation with p′ = O(p) satisfying
f ′(X) = hardmax(A(X))V (X).
The proof of Lemma A.2 is provided in Appendix F.
16
Start tokens.
Our technical proofs are occasionally simplified by including a “dummy token” whose value
is passed in self-attention layers as a default or null value. For example, in the proof of Lemma D.2, the
dummy token handles the case where the reference token does not appear previously in the sequence. While
we believe that this extra token is not necessary for our technical arguments, we include it for the sake of
simplicity.
We model this dummy token as a start-of-sequence token X0. Concretely, if we employ X0 in a self-
attention f ∈AttnN
m which takes as input X, we instead treat f as an attention unit in AttnN+1
m
that operates
on (X0, X1, . . . , XN). We assume that X0 is constant-valued, and therefore never both to pay attention to its
outputs; it’s only relevance is via its key and value embeddings K0(X0), V0(X0) ∈Rm. If X0 is unmentioned,
we assume that it does not exist, or is set such that its key embedding inner products are all zero.
Supplemental chain-of-thought tokens.
We periodically (see Theorem B.3 and the proofs of Corollar-
ies 3.5 and 4.3) consider transformers with supplemental blank “chain-of-thought” tokens appended to the
end of the sequence. Unlike the start token, these are only constant at initialization and may be used deeper
in the model to perform meaningful computations.
Let TransformerN,M
m,L,H,din,dout denote transformers with M −N extra blank elements appended to the input
sequence. Concretely, we represent T ∈TransformerN,M
m,L,H,din,dout as some T ′ ∈TransformerM
m,L,H,din,dout and
define the output T(X) for X ∈RN×din by letting Y ∈RM×din for Y1:N = X and YN+1:M = ⃗0, and letting
T(X) = T ′(Y ).
B
Proofs from Section 3.1
B.1
Proof of Lemma 3.2
Lemma 3.2. For any β, s, N ∈N, there exists a transformer routeβ,s ∈TransformerN
m,1,1 with m =
O(s4β log N) satisfying routeβ,s(Sent) = Rcvd for any valid (β, s)-routing (Sent, Rcvd).
The proof relies on a sparse propagation sequential primitive, which complements the sparse averaging
primitive of Sanford et al. (2023). For any Q ≤d, N, on input X = (X1, . . . , XN) ∈RN×d with Xi =
(zi, Si) ∈Rd−Q × [N]Q and bi = |{Sj ∋i : j ∈[N]}| ≤Q, we define
sparsePropagateQ,d(X)i =
(
1
bi
P
Sj∋i zj
if bi > 0,
0
otherwise.
Closely following the argument of Sanford et al. (2023), we show in Proposition B.1 that there is a self-
attention unit with embedding dimension m = max(d, O(q log N)) that computes sparsePropagateQ,d. This
construction is a key component of the single-layer transformer used in the proof of Lemma 3.2.
Proposition B.1. For any b ≤N and d, there exists a self-attention unit sparsePropagateQ,d ∈AttnN
m,p for
m = d+O(Q log N) and p = O(log N), which, given any input X with Xi = (zi, Si,⃗0) ∈Rd×
[N]
≤Q
×{0}m−Q−d
such that bi = |{Sj ∋i : j ∈[N]}| ≤Q for all i, has output sparsePropagateQ,d(X) satisfying
sparsePropagateQ,d(X)i = 1
bi
X
Sj∋i
zj.
The proof of Proposition B.1 appears in Appendix F.
Proof of Lemma 3.2. We construct a single-layer single-headed transformer with query, key, and value
embeddings Q, K, V and output MLP ψ. Q, K, V can be decomposed as Q = Q′ ◦ϕ, K = K′ ◦ϕ, V = V ′ ◦ϕ,
for some input MLP ϕ and embeddings Q′, K′, V ′. We fix Q′, K′, V ′ to be the respective embeddings of the self-
attention unit with embedding dimension m from Proposition B.1 that computes Y = sparsePropagates,m(X)
for XSrc = (zSrc, SSrc) for every Src ∈[N] to be determined. Hence, the proof entails designing element-
wise encoders ϕ = (ϕ1, . . . , ϕN) and decoders ψ = (ψ1, . . . , ψN) that compute Rcvd from Sent, using
17
sparsePropagate
2Y2 = 2 · sparsePropagate((z, S))2
2¯α
2Src
2Dest
Y1 = (0, 2)
ϕ
g
Dest
g
Msg
g
Src
˜α
z1
1
0
1
0
1
0
1
0
0
2
2
1
1
1
1
hey
yo
hey
yo
S1 = (0, 2)
S3 = (2, 4)
ϕ
Dest
Msg
yes
no
2
4
Sent3
g
Dest
g
Msg
g
Src
˜α
z3
0
1
0
1
1
0
3
3
3
3
2
2
4
4
yes
yes
no
no
Dest
Msg
hey
yo
0
2
Sent1
2Msg
= z1 + z3
1
1
2
1
2
0
1
hey
yes
asd
no
wyz
yo
1
3
3
1
0
2
2
4
Src
Msg
yo
yes
0
2
Rcvd2
ψ
1
Sent
ϕ
(z, S)
Y
Rcvd
ψ
Machine 0
Machine 4
Machine 1
Machine 3
Machine 2
Figure 3: A visualization of the construction used to prove Lemma 3.2 in three phases—the encoding of
each input SentSrc as embedding zSrc and subset SSrc with ϕ; the combination of those embeddings into
YDest via the simulation of sparsePropagates,m((z, S)); and the decoding of each YDest into output RcvdDest
with ψ. The figure provides an example of the encoding and decoding where machines 1 and 3 transmit
messages to machine 2. “Multiple hashing” is used to compute z1 and z3 by encoding each message in multiple
fixed-location “packets” in embedding space space. This redundancy ensures the possibility of machine 2
decoding Rcvd2 from Y2, due to each message occurring alone at least once in the encoding.
sparsePropagates,m as an intermediate step. A high-level overview of the proof construction is visualized in
Figure 3.
On input SentSrc, we use the encodings QSrc, KSrc, VSrc to specify that all tokens Dest with Dest ∈SentSrc
(or equivalently, all Dest with Src ∈RcvdDest) should receive a copy of the encoding of SentSrc. That is, we
set SSrc := {Dest ∈SentSrc} for each Src ∈[N]. This ensures that Y satisfies
YDest =
1
|RcvdDest|
X
Src∈RcvdDest
zSrc.
While it’s tempting to simply set each zSrc ∈Rm equal to a (βs)-dimensional vectorization of SentSrc, it
is unclear how to extract RcvdDest from each YDest, since each average performed by sparsePropagates,m will
combine multiple vector embeddings in a shared space. In order to avoid these troubles, we employ a multiple
hasing-based encoding that treats messages as “packets” identified by a message, a source, a destination, and
a “validity token” that can be used to determine whether a message is uncorrupted. We include multiple
18
copies of each packet in the encoding zSrc. For notational ease, we represent each zSrc ∈Rm as a collection of
packets
zSrc = (g
MsgSrc,j, g
SrcSrc,j, ]
DestSrc,j, αSrc,j)j∈[m′] ∈(Zβ
2p × [N] × [N] × {0, 1})m′,
where m = m′(3 + β).
To sparsely and redundantly encode each SentSrc as zSrc, we encode outgoing messages as packets
by utilizing the matrix A guaranteed by the following fact (which we use with n := N 2, b := s2, and
m′ := d = O(s4 log N)).
Fact B.2. For any n, b ≤n, and d ≥
12b2 ln n
, there exists a binary matrix A ∈{0, 1}n×d such that, for
every subset S ⊆[n] with |S| ≤b, the columns of the sub-matrix AS ∈{0, 1}|S|×d contains all S-dimensional
elementary vectors, i.e.,
e1, . . . , e|S|
is a subset of the columns of AS.
The proof of Fact B.2 is at the end of the section. We use the following rule to determine which (if any)
message to encode as a packet at each Src ∈[N] and j ∈[m′]. We let A(Src,Dest),j = AN(Src−1)+Dest,j for
notational convenience.
zSrc,j =
(Msg, Src, Dest, 1)
if (Msg, Dest) ∈SentSrc and A(Src,Dest),j = 1
and A(Src,Dest′),j = 0, ∀Dest′ ∈SentSrc \ {Dest} ,
(⃗0, 0, 0, 0)
otherwise.
In Figure 3, this encoding is visualized in the tables of “Machine 1” and “Machine 3,” where the entirety of
each message is encoded in two fixed and distinct locations in the embeddings z1 and z3, alongside metadata
about the source of message and the validity ˜α. Each message is encoded as multiple identical packets in
different embedding dimensions and a large fraction of embedding locations are left blank. These features are
critical for the proper evaluation of the decoding step ψ.
We analyze the Y = sparsePropagateβ,m(X) outputs, letting
YDest = (YDest,1, . . . , YDest,m′),
YDest,j ∈(Rβ × R × R × R)m′,
with all numbers represented with p-bit fixed precision. This analysis shows that there exists an element-wise
decoder MLP ψ satisfying ψDest(YDest) = RcvdDest for all Dest ∈[N]. For any j ∈[m′], observe from the
definition of zSrc and sparsePropagates,m that
YDest,j =:
MsgDest,j, SrcDest,j, DestDest,j, ¯αDest,j
=
1
|RcvdDest|
X
Src∈RcvdDest
g
MsgSrc,j, g
SrcSrc,j, ]
DestSrc,j, αSrc,j
.
Before formally analyzing this construction, we motivate its utility with Figure 3. The encoding 2Y2 of
Machine 2 contains four “clean” rows j with 2¯α2,j = 1, two “corrupted” rows with 2¯α2,j = 2, and one “blank”
row with 2¯α2,j = 0.
• The blank row contains no information about any incoming messages, since neither Machine 1 nor
Machine 3 encoded messages as packets in these locations. The fact that 2¯α2,j = 0 certifies the blankness
of this row, and hence, the decoder ψ can ignore it.
• The corrupted rows correspond to locations where both Machine 1 and Machine 3 saved messages as
packets. As a result, the corresponding embedding Y2,j = 1
2(z1,j + z3,j) is an average of two non-zero
embeddings and is hence “corrupted.” Because 2¯α2,j = 2, the decoder ψ detects the corruption and
ignores it when computing Rcvd2.
• The clean rows are locations where exactly one of Machine 1 and Machine 3 encoded a message. Hence,
these messages can be cleanly understood by the decoder ψ, which simply validates the “cleanliness” of
the row with 2¯α2,j = 1, determines whether Machine 2 is indeed the target recipient of the respective
message, and saves all such messages in the decoding Rcvd2.
19
We prove the validity of this intuition by ensuring that the encoding scheme successfully encodes each
incoming message in a clean row and that the category of each row (blank, corrupted, or clean) can be
detected by the decoder ψ. We observe the following sequence of facts about every YDest. Let
RelevantDest := {(Msg, Src′, Dest′) : Src′ ∈RcvdDest, (Msg, Dest′) ∈SentSrc′}
denote the set of all messages sent by sources of messages sent to Dest.
1. Consider any outgoing message (Msg, Src′, Dest′) ∈RelevantDest. By the property of A guaranteed by
Fact B.2, there exists some j such that A(Src′,Dest′),j = 1 and A(Src′′,Dest′′),j = 0 for every (Src′′, Dest′′) ∈
RelevantDest \ {(Src′, Dest′)} . As a result of the definition of the encoding z and the averaged
representation of YDest:
YDest,j =
1
|RcvdDest| (Msg, Src′, Dest′, 1) .
(1)
2. Conversely, if ¯αDest,j = 1/|RcvdDest|, then there exists a unique (Msg, Src′, Dest′) ∈RelevantDest such
that (1) is satisfied.
3. If at least one message is received, then the minimal nonzero value of ¯αDest is 1/|RcvdDest|.
We design ψDest to uniquely identify RcvdDest from YDest as follows. If at least one message is received,
then 1/|RcvdDest| can be identified by finding the smallest nonzero value of ¯αDest. The decoder ψ inspects
every YDest,j satisfying ¯αDest,j = 1/|RcvdDest|, which therefore satisfies
|RcvdDest| · (MsgDest,j, SrcDest,j, DestDest,j) ∈RelevantDest.
Thus, if |RcvdDest| · DestDest,j = Dest, then |RcvdDest| · (MsgDest,j, SrcDest,j) ∈RcvdDest, and ψ encodes it as
such.
Fact B.2. For any n, b ≤n, and d ≥
12b2 ln n
, there exists a binary matrix A ∈{0, 1}n×d such that, for
every subset S ⊆[n] with |S| ≤b, the columns of the sub-matrix AS ∈{0, 1}|S|×d contains all S-dimensional
elementary vectors, i.e.,
e1, . . . , e|S|
is a subset of the columns of AS.
Proof. Let col(A) denote the set of columns of A. We use the probabilistic method and consider A with iid
entries Ai,j ∼Bernoulli(
1
b+1). We bound the probability of failure:
Pr
∃S ∈
[n]
≤b
s.t.
e1, . . . , e|S|
̸⊂col(AS)
≤b · nb Pr [ei ̸∈col(AS)]
≤nb+1
1 −
1
b + 1 ·
1 −
1
b + 1
b!d
≤nb+1
1 −
1
e(b + 1)
d
≤nb+1 · exp
−
d
e(b + 1)
< exp
(b + 1) ln n −
d
3(b + 1)
≤1.
Therefore, there exists a matrix A with the claimed property.
20
Attention head
Attention head
Sparse propagation
Local
Softmax attention
Attention head
Local
Q(·)
K(·)⊤
V (·)
Multiple
hashing
Figure 4: To simulate MPC, the local computation within each machine is pushed inside Q(·), K(·), V (·), and
then the pairwise attention matrix performs message routing. To ensure proper routing and also that the
outputs of Q(·), K(·), V (·) are all tall-and-skinny matrices, the construction carefully utilizes both multiple
hashing and sparse propagation.
B.2
Proof of Theorem 3.1
We give a generalization of Theorem 3.1 that simulates a broader family of MPC protocol, including those
with more than n machines (i.e. γ ≥δ). We accommodate this generalization by simulating MPC protocols
with the generalized transformer family TransformerN,M
m,L,H detailed in Appendix A with supplemental blank
“chain-of-thought” tokens.
Theorem B.3 (Generalization of Theorem 3.1). For constant γ, δ > 0 and any potentially randomized
R-round (γ, δ)-MPC protocol π on nin input words and nout ≤nin output words, there exists a transformer T ∈
TransformerN,M
m,L,H with N = nin, M = max(nin, O(n1+γ−δ
in
)), m = O(n4δ
in log nin), L = R+1, H = O(log log nin)
such that
T(Input):nout = π(Input).
Theorem 3.1 is an immediate consequence of Theorem B.3 by noting that M = N for sufficiently large
nin when γ < δ. Its central construction is summarized in Figure 4.
Proof. Consider any MPC protocol π with q = O(n1+γ−δ
in
) machines and s = O(nδ
in) local memory
that, following the notation of Definition 2.3, maps Input ∈Znin
2p
to Output ∈Znout
2p
with interme-
diates MachineIn(1), . . . MachineIn(R) and MachineOut(1), . . . , MachineOut(R) and deterministic functions
(Localr,i)r∈[R],i∈[q] with
MachineOut(r)
i
= Localr,i(MachineIn(r)
i ).
To simulate the protocol, we let every machine i ∈[q] correspond to a particular position in the transformer’s
context. A transformer that simulates π can then be constructed that consolidates Input onto ⌈nin/s⌉machines
to match MachineIn(1); computes MachineIn(r+1) from MachineIn(r) for each r = 1, . . . , R−1; and computes
and properly distributes Output from MachineIn(r). These three elements of the construction exist due to
the following lemmas, which are proved later.
21
Lemma B.4. For any MPC protocol π with local memory s and q machines with nin-word inputs, there
exists a transformer init ∈Transformernin,max(nin,q)
s,1,1,din,dout
with din = 1 and dout = s, which, given Input ∈Zn
2p, has
output satisfying init(Input) = MachineIn(1).
Lemma B.5. For any R-round MPC protocol π with local memory s and q machines and any r ∈[R −1],
there exists a transformer round(r) ∈Transformerq
m,1,H,din,dout with H = O(log log q), m = O(s4 log q), and
din = dout = s which, given any valid input X = MachineIn(r) ∈Zq×m
2p
under the MPC protocol in vectorized
form, has output satisfying round(r)(X) = MachineIn(r+1).
Lemma B.6. For any R-round MPC protocol π with local memory s and q machines with nout-word output,
there exists a transformer final ∈Transformerq,max(nout,q)
s,1,1,din,dout
for din = s and dout = 1, which, given input
X = MachineIn(R), has output final(X) with final(X)i,1 = Outputi ∈Z2p.
The proof immediate from the three lemmas. We construct the final transformer T by stacking the
single-layer constructions as a single transformer with embedding dimension m:
T = final ◦round(R−1) ◦· · · ◦round(1) ◦init.
The proofs of Lemmas B.4 and B.6 rely on simple constructions with fixed attention matrices and appear
in Appendix F. The proof of Lemma B.5 relies on Lemma 3.2 and is proved in the following section.
Proof of round(r) construction.
To prove the existence single-layer transformer that simulates round(r),
we separate the computational task into two steps: (i) obtaining MachineOut(r) from MachineIn(r) and
(ii) obtaining MachineIn(r+1) from MachineOut(r). Because the former requires no communication between
machines, we can encode that conversion in the input MLP to the transformer.
The nontrivial part of the reduction is thus the latter step, which we obtain by utilizing multiple single-
headed attention units routeβ,s of Lemma 3.2 to route messages of different sizes to their recipients. The
difficulty in this task is the mismatch in functionality between the two computational models: while the MPC
model ensures that each recipient automatically receives its intended messages, transformers must implement
this functionality manually, while ensuring that multiple messages do not overwrite one another.
The following lemma implements that routing functionality for all messages, using different attention
heads depending on the size of the message. We prove Lemma B.5 at the end of the section as a simple
modification of Lemma B.7.
Lemma B.7. For any R-round MPC protocol π with local memory s and q machines and any r ∈[R −1],
there exists a transformer route(r) ∈Transformerq
m,1,H with H = O(log log q) and m = O(s4 log q), which,
given any valid input X = MachineOut(r) ∈Zq×m
2p
under the MPC protocol in vectorized form, has output
satisfying route(r)(X) = MachineIn(r+1).
Because at most s messages can be shared and received by each machine, and each message is of size at
most s, we can prove an single-headed alternative to Lemma B.7 with a somewhat suboptimal dependence on
embedding dimension. By applying by Lemma 3.2 with message size β = s, bounded number of messages s,
and context length N = q, there exists a transformer routes,s with H = 1 and m = O(s5 log q) that computes
MachineIn(r+1) from MachineOut(r+1) by regarding each outgoing message as belonging to Zs
2p by adding
padding dimensions as needed.
We improve the embedding dimension to m = O(s4 log q) by running in parallel O(log log N) transformers
guaranteed by Lemma 3.2 that encode differently sized messages. The number of heads H increases at a
doubly-logarithmic rate because of a doubling trick employed on the size of message encodings used by
constituent part.
22
Proof. We describe an implementation of route(r) by considering any fixed input MachineOut(r) ∈Zq×m
2p
. For
each i ∈[q] and some integer sequence 1 = β0 < β1 < · · · < βH = s + 1, we partition MachineOut(r)
i
into H
disjoint subsets as follows. For any h ∈[H], let
Senth
i :=
n
(Msg, Dest) ∈MachineOut(r)
i
: dim(Msg) ∈[βh−1, βh]
o
,
Rcvdh
i :=
n
(Msg, Src) ∈MachineIn(r+1)
i
: dim(Msg) ∈[βh−1, βh]
o
,
and note that MachineOut(r)
i
= ˙SH
h=1Senth
i and MachineIn(r+1)
i
= ˙SH
h=1Rcvdh
i .
For each h ∈[H], note that dim(Msg) ≤βh, and
Senth
i
=
Rcvdh
i
≤s/βh−1. As a result, Lemma 3.2
guarantees the existence of a single-headed transformer route(r)
h
such that route(r)
h (Senth) = Rcvdh) with
embedding dimension mh ≤Cs4βh log(q)/β4
h−1 for some sufficiently large universal constant C.
We defined route(r) as the computation of route(r)
1 , . . . , route(r)
H as H parallel heads of self-attention with
disjoint embeddings concatenated into in m-dimensional embedding space with m = PH
h=1 mh. We conclude
by letting
βh =
(
1
if h = 0,
min(2β3
h−1, q + 1)
if h ∈[H],
noting that βH = q + 1 for H = O(log log q), and bounding m:
m ≤
H
X
h=1
Cs4 log(q)βh
β4
h−1
≤2Cs4 log(q) ·
H
X
h=1
1
βh−1
≤2Cs4 log(q) ·
H
X
h=1
1
2h−1 = O(s4 log q).
Proof of Lemma B.5. To simulate a round of MPC protocol π by mapping MachineIn(r) and ρr to MachineIn(r+1),
the single-layer transformer round(r) first computes MachineOut(r) element-wise and then properly routes
messages in MachineOut(r) to their proper destination. We can define round(r) = route(r) ◦Localr for route(r)
in Lemma B.7 and Localr,i(MachineIn(r)
i , ρr,i) = MachineOut(r)
i . This can be immediately constructed as a
single-layer transformer by prepending the embeddings Q, K, V of the construction of route(r) with Localr,
using Q ◦Localr, K ◦Localr, V ◦Localr as the embeddings of round(r).
B.3
Additional graph problems solvable by log-depth transformers
Theorem 8.1 and Corollary 8.2 of Coy and Czumaj (2022) give efficient MPC protocols for other graph problems
besides connectivity, and therefore, as corollaries of Theorem 3.1, we also obtain log-depth transformers for
these problems.
Corollary B.8 (Spanning forest construction). For any constant ϵ ∈(0, 1) and any D ≤N, there exists
a transformer in TransformerN
m,L,H with m = O(N ϵ), H = O(log log N), and L = O(log D) that computes a
rooted spanning forest of any input graph G = (V, E) with |V |, |E| = O(N) where each connected component
has diameter at most D.
Corollary B.9 (Minimum spanning forest construction). For any constant ϵ ∈(0, 1) and any DMSF ≤N,
there exists a transformer in TransformerN
m,L,H with m = O(N ϵ), H = O(log log N), and L = O(log DMSF )
that identifies the connected components of any input graph G = (V, E) with |V |, |E| = O(N) and poly(N)-
bounded integer weights whose minimum spanning forest has diameter at most DMSF .
23
(Qh,i, Kh,i, Vh,i)
Token machines
Key/value
propagation
machines
Query
propagation
machines
Q⊤
h,iKh,i′
Inner product machines
Figure 5: This construction employs M 2 inner product machines to compute the entries of the softmax
matrix, and M token machines to compute all values of Q(·), K(·), V (·). What is most complex about the
construction are the additional machines and message routing needed to propagate these values efficiently
between the inner product machines and the token machines, in particular carefully aggregating the output of
the attention mechanism and computing its normalization. To this end, the protocol uses additional machines,
organized into a tree with branching factor b = O(N δ′−δ) and depth D = O( 1+α
δ′−δ).
C
Proofs from Section 3.2
C.1
Proof of Theorem 3.4
As in Appendix B.2, we give and prove a generalized version of Theorem 3.4 that broadens the family of
considered transformers to include masked models and those that contain extra blank chain-of-thought tokens,
using notation from Appendix A.
Theorem C.1 (Generalization of Theorem 3.4). For any transformer T ∈TransformerN,M
m,L,H (or MaskTransformerN,M
m,L,H)
with mH = O(N δ) for δ ∈(0, 1) and M = Θ(N 1+α) for α ≥0 and for any δ′ ∈(δ, 1), there exists an
O( L(1+α)
δ′−δ )-round (1 + 2α + δ′, δ′)-MPC protocol with q = O(M 2) machines with s = O(N δ′) local memory
that outputs the same sequence as T(X) for all X ∈RN.
Theorem 3.4 is an immediate consequence by setting M := N and α := 0.
Proof. It suffices to show that an O( 1+α
δ′−δ)-round MPC protocol π that simulates a single-layer transformer
T ∈TransformerM
m,m,m,1,H with m-dimensional input and output embeddings since a depth-L transformer
can be constructed by applying L such protocols sequentially. Moreover, we can ignore the difference between
the input context length N and the context length with padding M by assuming that the input contains M
tokens.
Concretely, we consider H heads with embeddings (Qh, Kh, Vh)h∈[H], element-wise output MLP ψ =
(ψ1, . . . , ψM), and any fixed masks Λ1, . . . , ΛH ∈{−∞, 0}M×M. We show that there exists some π such that
24
for any Input = X ∈RM×m,
π(X) = ψ
X +
H
X
h=1
softmax(Qh(X)Kh(X)T + Λh)Vh(X)
!
,
where numbers in X and all intermediate products of the transformer computation can be represented with
p = O(log M) bit precision.
Our MPC protocol π, which will use q = O(M 2) machines and s = Θ(N δ′) words of local memory per
machine, assigns each of the q machines to one of four possible roles: token machine, inner product machine,
query propagation machine, and key/value propagation machine. We describe these machines below. For the
sake of readability, we identify machines with easily interpretable descriptions and use the bijection ID to map
each of those to a token in [q] that is used for routing messages. Our protocol has two important parameters:
b = ⌊s/(4mH)⌋= O(N δ′−δ) is the branching factor of the protocol, and D = ⌈logb(M)⌉= O( 1+α
δ′−δ) is the
depth of the protocol.
At a high level (see Figure 5 for a corresponding diagram), the protocol involves computing all intermediate
products of the of a transformer unit by performing MLP computations in N token machines, computing
inner products in N 2 inner product machines, and using O(N 2) other propagation machines arranged in trees
to share information between the two in O(D) rounds. The protocol draws inspiration from Appendix C.6.1
of Sanford et al. (2023), which uses a similar construction to simulate transformers with Congest protocols
on fixed graphs. It is also similar to the MPC implementation of the MPI AllReduce functionality (MPICH,
2023) described by Agarwal et al. (2014).
• Machine i ∈[M] is a token machine that performs all element-wise computation on the ith token
embedding, including the computation of (Qh,i(Xi), Kh,i(Xi), Vh,i(Xi))h∈[H] and the final MLP output
ψi. Let ID(i) = i.
• Machine (i, i′) ∈[M]2 is an inner product machine designed to compute the inner products (Qh,i(Xi)TKh,i′(Xi′))h∈[H].
• Machine (Q, i, d, k) for token i ∈[M], depth d ∈[D −1] and position k ∈[bd] is a query propagation
machine. This machine is responsible for handling communication of query tokens (Qh,i(Xi))h∈[H] and
of all partially-computed attention outputs for the ith token between token machine i and inner product
machines (i, i′) for
i′ ∈Descendantsd,k :=
bD−d(k −1), . . . , bD−dk
∩[M].
Concretely, if ℓ= 1, then the machine communicates with token machine i and query propagation
machines (Q, i, d + 1, k′) for
k′ ∈Childrenk := {b(k −1) + 1, . . . , bk} .
If ℓ= D −1, then it communicates with inner product machines (i, i′) for i′ ∈Childrenk ∩[M] and
query propagation machine (Q, i, d −1, ⌊k/b⌋). Otherwise, it communicates with query propagation
machines (Q, i, d −1, Parentk), for Parentk := ⌊k/b⌋, and (Q, i, d + 1, k′) for k′ ∈Childrenk.
• Machine (KV, i, d, k) is a key/value propagation machine.
This machine is analogous to a query
propagation machine, except that it is responsible for the communication of key and value to-
kens (Qh,i(Xi), Vh,i(Xi))h∈[H] between token machine i and inner product machines (i, i′) for i′ ∈
Descendantsd,k.
Since the total number of machines is q = M + M 2 + M PD−1
d=1 bd = O(M 2), we conclude that the global
memory of the protocol is qs = O(N 2+2α+δ′), which means the protocol is (1 + 2α + δ′, δ′)-MPC. We simulate
the transformer using a four stage protocol using 2D + 3 = O( 1+α
δ′−δ) rounds of MPC computation.
25
Stage 1: Token dispersion.
Because the input to an MPC protocol Input = X is divided equally
among machines 1, . . . , ⌈MmH/s⌉, the first round of MPC computation routes each input token Xi to its
respective token machine. This is completed by setting (i, Xi) ∈MachineOut(1)
i′
if (i, Xi) ∈MachineIn(1)
i′ .
Thus, MachineIn(2)
i
= {(Src, Xi)} for all token machines i ∈[M].
Stage 2: Embedding propagation.
In rounds 2, . . . , D + 1, π computes the respective key, query, and
value embeddings in each token machine and propagate them to respective inner product machines using the
query and key/value propagation machines. Concretely:
• In round 2, each token machine i (whose memory contains Xi) computes m-dimensional embeddings
embeddings Qi := (Qh,i(Xi))h∈[H], Ki := (Kh,i(Xi))h∈[H], Vi := (Vh,i(Xi))h∈[H]. It transmits each
embedding to the respective depth-1 query and key/value propagation machine nodes, while also
preserving knowledge of its own Xi. (In all further rounds, we assume that ((i, Xi)) ∈MachineOut(r)
i
to ensure that token machine i can compute the skip-level connection at the end.) That is,
MachineOut(2)
i
= {(i, Xi)}
∪{(ID(Q, i, 1, k′), Qi) : k′ ∈Children1}
∪{(ID(KV, i, 1, k′), (Ki, Vi)) : k′ ∈Children1} .
Note that the total amount of messages sent is b · mH + 2b · mH + m ≤s and that the only machines
receiving messages are size m-messages by token machines and size ≤4mH messages by query and
key/value propagation machines.
• In rounds r ∈{3, . . . , D}, each query and key/value propagation machine of depth d = r −2 passes
embeddings onto their successors. That is,
MachineOut(r)
ID(Q,i,d,k) = {(ID(Q, i, d + 1, k′), Qi) : k′ ∈Childrenk} ,
MachineOut(r)
ID(KV,i,d,k) = {(ID(KV, i, d + 1, k′), (Ki, Vi)) : k′ ∈Childrenk} .
• In round D + 1, the depth-(D −1) query and key/value propagation machines pass their embeddings
onto their respective inner product machines. That is,
MachineOut(D+1)
ID(Q,i,D−1,k) = {(ID(i, k′), Qi) : k′ ∈Childrenk ∩[M]} ,
MachineOut(D+1)
ID(KV,i,D−1,k) = {(ID(k′, i), (Ki, Vi)) : k′ ∈k′ ∈Childrenk ∩[M]} .
Stage 3: Softmax computation.
In rounds D+2, . . . , 2D+2, computes each inner product and iteratively
builds up each attention output by accumulating partial softmax computations. For each query propagation
machine (Q, i, d, k) and h ∈[H], we let Si,d,k,h and Zi,d,k,h denote its partial normalization and softmax
computations respectively. That is,
Zi,d,k,h =
X
i′∈Descendantsd,k
exp(Qh,i(Xi)TKh,i′(Xi′))1 {Λi,i′ = 0}
=
(P
k′∈Childrenk Zi,d+1,k′,h
if d ≤D −1,
exp(Qh,i(Xi)TKh,k(Xk))1 {Λi,k = 0}
if d = D.
Si,d,k,h =
1
Zi,d,k,h
X
i′∈Descendantsd,k
exp(Qh,i(Xi)TKh,i′(Xi′))Vh,i′(Xi′)1 {Λi,i′ = 0}
=
(P
k′∈Childrenk
Zi,d+1,k′,h
Zi,d,k,h
· Si,d+1,k′,h
if d ≤D −1,
Vh,k(Xk)1 {Λi,k = 0}
if d = D;
Note that Si,0,1,h = (softmax(Qh(X)Kh(X)T + Λh)Vh(X))i and let Si,d,k = (Si,d,k,h)h∈[H] ∈RH×m and
Zi,d,k = (Zi,d,k,h)h∈[H] ∈RH
26
• In round D + 2, each inner product machine computes its respective inner products and passes its
partial softmax computations to its parent query propagation machine. As a result of round D + 1,
each inner product machine (i, i′) recently received the embeddings necessary to compute the relevant
inner product:
MachineIn(d+2)
ID(i,i′) = {(ID(Q, i, D −1, Parenti), Qi), (ID(KV, i′, D −1, Parenti′), (Ki′, Vi′))} .
It propagates the respective partial computations Si,D,i′ and Zi,D,i′ as follows:
MachineOut(D+2)
ID(i,i′) = {(ID(Q, i, D −1, Parenti), (Si,D,i′, Zi,D,i′))} .
Note that each depth-(D−1) query propagation machine receives messages of size at most b·(m+1)H ≤s.
• In rounds r ∈{D + 3, . . . , 2D}, partial softmax computations are received by query propagation
machines of depth d = 2D + 1 −r, added together, and passed along to their parent machines. That is,
given
MachineIn(r)
ID(Q,i,d,k) = {(ID(Q, i, d + 1, k′), (Si,d+1,k′, Zi,d+1,k′)) : k′ ∈Childrenk} ,
each respective machine computes Si,d,k and Zi,d,k recursively and propagates
MachineOut(r)
ID(Q,i,d,k) = {(ID(Q, i, d −1, Parentk), (Si,d,k, Zi,d,k)} .
• In round 2D + 1, the top-most query propagation tokens pass their partial sums to the token machines:
MachineOut(2D+1)
ID(Q,i,1,k) = {(i, (Si,1,k, Zi,1,k))} .
• In round 2D + 2, the token machines compute their respective output of the transformer, T(X)i. Given
input
MachineIn(2D+2)
i
= {(k′, (Si,1,k′, Zi,1,k′)) : k′ ∈Children1} ∪{(i, Xi)} ,
the token machine i computes Si,0,1 and Hi,0,1 and then
T(X)i = ψi
Xi +
H
X
h=1
softmax(Qh(X)Kh(X)T + Λh)T
i Vh(X)
!
= ψi
Xi +
H
X
h=1
Si,0,1,h
!
.
This quantity is used as an intermediate product for the final phase of computation.
Stage 4: Token compression.
We invert Stage 1 by properly compressing the MPC output in the final
round 2D + 3. That is, we let MachineOut(2D+2)
i
= {(⌊imH/s⌋+ 1, T(X)i)} for each token machine i ∈[M],
which ensures that the outputs are condensed in the proper order in machines 1, . . . , ⌈MmH/s⌉.
Precision analysis.
In order for the proof to be fully sound, care must be taken to ensure that the
computation of each self-attention output Si,0,1,h is handled with proper numeric precision, as discussed in
Appendix A. We show that each Si,0,1,h is a valid implementation of its corresponding self-attention unit, per
Definition A.1.
To do so, we let ˆSi,d,k,h and ˆZi,d,k,h denote the p-bit representations of Si,d,k,h and Zi,d,k,h, where scalars
of ˆSi,d,k,h and log( ˆZi,d,k,h) are represented as discretized rational numbers z satisfying |z| ≤1
22p/2 and
z · 2p/2 ∈Z. For some sufficiently small p′ = Θ(p), we assume that all embeddings Qh(X), Kh(X), Vh(X)
have scalars z satisfying |z| ≤1
22p′/2 and z · 2p′/2 ∈Z. We prove that for each h ∈[H],
Si,0,1,h −ˆSi,d,k,h
∞= O
1
2p′
.
27
Boundedness of intermediate representations is not an issue because
log(Zi,d,k,h) ≤O(log(N) + max
i,i′ |Q(X)T
i K(X)i′|) = exp(O(p′)),
and
∥Si,d,k,h∥∞≤∥V (X)∥∞≤2p′/2.
It remains to show that that all intermediate representations are sufficiently close to their exact counterparts.
We prove the following via an inductive argument for d = D, D −1, . . . , 0:
log(Zi,d,k,h) −log( ˆZi,d,k,h)
≤(2b)D−d
2p/2
,
(2)
Si,d,k,h −ˆSi,d,k,h
∞≤2p′/2(8b)D−d
2p/2
.
(3)
If (3) holds for d = 0, then the claim holds for sufficiently large p = Θ(p′).
For the base case D, we verify (3) by
Si,D,k,h −ˆSi,D,k,h
∞=
Vh,k(Xk)1 {Λi,k = 0} −ˆSi,D,k,h
∞≤
1
2p/2 ,
due to the ability to access Vh,k(Xk) and round it directly. We verify (2) due to the immediate access to and
boundedness of Qh,i(Xi)TKh,k(Xk):
|log(Zi,d,k,h)| ≤
Qh,i(Xi)TKh,k(Xk)
≤∥Qh,i(Xi)∥2 ∥Kh,k(Xk)∥2 ≤N · 2p′/2.
We prove the inductive step for d −1, assuming that the inductive hypothesis holds for d. We first address
ˆZi,d−1,k,h by employing the Lipschitzness of the log-sum-exp function.
log(Zi,d−1,k,h) −log( ˆZi,d−1,k,h)
≤
1
2p/2 +
log
X
k′
exp(log(Zi,d,k′,h))
!
−log
X
k′
exp(log( ˆZi,d,k′,h))
!
≤
1
2p/2 +
X
k′
log(Zi,d,k′,h) −log( ˆZi,d,k′,h)
≤
1
2p/2 + b · (2b)D−d
2p/2
≤(2b)D−d+1
2p/2
.
To obtain (3) for d −1, we first note that for sufficiently large p:
1 −
ˆZi,d,k′,hZi,d−1,k′,h
Zi,d,k,h ˆZi,d−1,k′,h
=
1 −exp
log
ˆZi,d,k′,h
Zi,d,k′,h
!
+ log
Zi,d−1,k,h
ˆZi,d−1,k,h
!!
≤1 + 2
log
ˆZi,d,k′,h
Zi,d,k′,h
+
log Zi,d−1,k,h
ˆZi,d−1,k,h
!
≤4 · (2b)D−d+1
2p/2
.
28
We conclude by using the fact that each Si,d−1,k,h is a convex combination of other Si,d,k,h.
Si,d−1,k,h −ˆSi,d−1,k,h
∞≤
1
2p/2 +
X
k′
Zi,d,k′,h
Zi,d−1,k′,h
Si,d,k′,h −
ˆZi,d,k′,h
ˆZi,d−1,k′,h
ˆSi,d,k′,h
∞
≤
1
2p/2 +
X
k′
Zi,d,k′,h
Zi,d−1,k′,h
Si,d,k′,h −
ˆZi,d,k′,hZi,d−1,k′,h
Zi,d,k,h ˆZi,d−1,k′,h
ˆSi,d,k′,h
∞
≤
1
2p/2 +
X
k′
Zi,d,k′,h
Zi,d−1,k′,h
Si,d,k′,h −ˆSi,d,k′,h
∞
+
X
k′
Zi,d,k′,h
Zi,d−1,k′,h
ˆSi,d,k′,h
∞
1 −
ˆZi,d,k′,hZi,d−1,k′,h
Zi,d,k,h ˆZi,d−1,k′,h
≤
1
2p/2 + 2p′/2(8b)D−d
2p/2
+ 2p′/2 X
k′
Zi,d,k′,h
Zi,d−1,k′,h
1 −
ˆZi,d,k′,hZi,d−1,k′,h
Zi,d,k,h ˆZi,d−1,k′,h
≤2 · 2p′/2(8b)D−d
2p/2
+ 2p′/2 · 4 · (2b)D−d+1
2p/2
≤2p′/2(8b)D−d+1
2p/2
.
Owing to the fact that D and p′ are constants and b = N O(1), a sufficiently large choice of p guarantees
that the implementation is valid.
C.2
Proof of Corollary 3.5
Corollary 3.5. Let ϵ ∈(0, 1) be any constant, and let D ≥N ϵ. Assume Conjecture 2.4, and suppose there
exists T ∈TransformerN
m,L,H with mH = O(D1−ϵ) that decides connectivity of any input graph with connected
components having diameter ≤D. Then L = Ω(log D).
We prove Corollary 3.5 by combining Theorem C.1 and Conjecture 2.4.
Proof. Fix any D ≤N with D ≥N ξ for some ξ ∈(0, 1]. Let C1 denote a cycle graph on D vertices, and let
C2 denote the union of two cycle graphs each with D/2 vertices.
Suppose there is a transformer T ∈TransformerN
m,L,H with mH = O(D1−ϵ) that determines the connectiv-
ity of graphs with at most N edges and connected components with diameter at most D. We will show that
it can be used to design an Θ(L)-round MPC protocol π that distinguishes graphs C1 and C2 with n = D
edges.
Let π′ be an MPC protocol that exactly computes the output of T using taking R = O(L) rounds with
local memory s = O(D1−ϵ/2) and q = O(N 2) machines, which is guaranteed to exist by Theorem C.1.
Let n := 2
D
4
and k :=
N
n
. We design π with the same local memory and machine count to determine
the identity of input graph G = (V, E) ∈{C1, C2} provided as an arbitrary sequence of n edges. Let u ∈V
be an arbitrary vertex in G.
Using a constant number of MPC rounds, π converts G into a graph G′ = (V ′, E′) with |E′| = kn+k ≤N
and diameter n + 2 ≤D such that G′ is connected if and only if G = C1. We do so by letting G′ be composed
of k copies G1, . . . , Gk of G on separate vertices, along with k extra edges connecting the vertex corresponding
to u in each Gj (say uj ∈Gj) to u1 ∈G1. This ensures that the connectivity correspondence and edge
count diameter bounds are met. Since G′ can be produced by simply copying edges from G and adding an
additional edge each time an edge containing u is copied, π can produce G′ in O(1) rounds.
Then, π simulates π′ on G′ and returns its output. Since G′ is connected if and only if G = C1, this
protocol suffices to distinguish C1 and C2. Because the protocol uses s = O(n1−ϵ/2) local memory and
q = O(n2/ξ) machines, Conjecture 2.4 implies that π (and hence T) only exists if L = Ω(log n) = Ω(log N).
29
D
Proofs from Section 4.1
D.1
Proof of Theorem 4.2
Theorem 4.2. For any k ∈N and alphabet Σ with |Σ| ≤N, there exists T ∈MaskTransformerN
m,L,H that
computes hopk : ΣN →(Σ ∪{⊥})N with m = O(1), L = ⌊log2 k⌋+ 2, and H = 1.
Proof. We design a masked transformer that implements hopk in two phases. The first two layers compute
find1
X(i) for each i ∈[N] using a similar approach to the induction heads construction of Bietti et al. (2023).
The subsequent layers employ a doubling trick to compute each find2ℓ−2
X
(i) after ℓlayers.
To do so we employ two technical lemmas (which are proved in Appendix F.4) that describe the
implementation of masked self-attention units that copy .
Lemma D.1. For some m ≥d + 2, τ : [N] × Rm →[N], and ρ : Rm →Rd, there exists an attention head
lookUpτ,ρ ∈MaskAttnN
m with precision p = O(log N) and m ≥d+2 satisfying lookUpτ,ρ(X)i,:d = ρ(Xτ(i,Xi)).
Lemma D.2. For any finite alphabet Σ, m ≥d + 2, µ1, µ2 : Rm →Σ, and ρ : Rm →Rd, there exists an
attention head lastOccurrenceµ,ρ ∈MaskAttnN
m with precision p = O(log(N |Σ|)) such that,
lastOccurrence(X)i,:d =
(
ρ(⃗0)
if ∀i′ < i : µ1(Xi′) ̸= µ2(Xi),
ρ(Xi′)
if i′ = max {i′ < i : µ1(Xi′) = µ2(Xi)} .
The first layer obtains the previous token Xi−1 from each Xi. This is accomplished via the self-attention
head lookUpτ,ρ with τ(i, Xi) = i −1 and ρ(Xi) = Xi.
The second layer retrieves (find1
X(i), Xfind1
X(i)) for each i ∈[N] by finding the most recent token whose
preceding token is Xi. It does so by employing the lastOccurrenceµ1,µ2,ρ primitive on the intermediate state
X1
i = (Xi, Xi−1) with µ1(X1
i ) = Xi−1, µ2(X1
i ) = Xi, and ρ(X1
i ) = (i, Xi).
• If find1
X(i) > 0, then lastOccurrenceµ1,µ2,ρ(X1
i ) = (find1
X(i), Xfind1
X(i)).
• Otherwise, it obtains ⃗0 and performs no further passing, returning ⊥after all L layers.
If k = 1, the transformer returns T(X)i = Xfind1
X(i) = hopk(X)i.
Otherwise, let k := P⌊log2 k⌋
j=0
kj2j for some kj ∈{0, 1}, and let k:ℓ= Pℓ
j=0 kj2j. Construct a transformer
inductively to ensure that the ith output of the ℓth layer Xℓ
i ∈Rm for ℓ≥2 contains an encoding of
Xi, find2ℓ−2
X
(i), Xfind2ℓ−2
X
(i), findk:ℓ−2
X
(i), Xfind
k:ℓ−2
X
(i)
.
Note that the base case holds for ℓ= 2, since findk:0
X (0) = find1
X(0) if k0 = 0 and is i otherwise.
For each ℓ= 1, . . . , ⌊log2 k⌋+ 1, we assume that the inductive hypothesis holds up to layer ℓand prove
that it also holds for layer ℓ+ 1. To do so, we use a lookUpτ,ρ self-attention head with τ(i, Xℓ
i ) = find2ℓ−2
X
(i)
and
ρ(Xℓ
i ) = (find2ℓ−2
X
(i), Xfind2ℓ−2
X
(i), findk:ℓ−2
X
(i), Xfind
k:ℓ−2
X
(i)),
which ensures that Xℓ+1
i
can encode
find2ℓ−1
X
(i) = find2ℓ−2
X
(find2ℓ−2
X
(i))
Xfind2ℓ−1
X
(i) = Xfind2ℓ−2
X
(find2ℓ−2
X
(i))
findk:ℓ−1
X
(i) =
(
findk:ℓ−2
X
(find2ℓ−2
X
(i))
if kℓ−1 = 1
findk:ℓ−2
X
(i)
if kℓ−1 = 0
Xfind
k:ℓ−1
X
(i) =
Xfind
k:ℓ−2
X
(find2ℓ−2
X
(i))
if kℓ−1 = 1
Xfind
k:ℓ−2
X
(i)
if kℓ−1 = 0.
30
As a result, the output of layer L = ⌊log2 k⌋+ 2 contains an encoding of
Xfind
k:L−2
X
(i) = Xfindk
X(i) = hopk(X)i
for each i ∈[N]. This is returned as the output of T(X).
D.2
Proof of Corollary 4.3
Corollary 4.3. Assuming Conjecture 2.4, for any constants ξ ∈(0, 1/2] and ϵ ∈(0, 1), and any even
k = Θ(N ξ), every transformer T ∈MaskTransformerN
m,L,H with mH = O(k1−ϵ) that computes hopk has depth
L = Ω(log k).
Proof. The proof is analogous to that of Corollary 3.5. Let C1 be a cycle on k vertices, and C2 be the
union of two cycles each on k/2 vertices. So both C1 and C2 have k edges. We show that the existence of
T ∈TransformerN
m,L,H with mH = O(k1−ϵ) such that T(X) = hopk(X) can be used to design an Θ(L)-round
MPC protocol π to solve the task.
As a result of Theorem C.1, there exists an MPC protocol π′ that exactly computes T with R = Θ(L)
rounds with local memory s = O(D1−ϵ/2) and q = O(N 2) machines. On input G = (V, E) ∈{C1, C2}, we
design a constant-round protocol that computes an sequence X ∈ΣN such that hopk(X)N exactly determines
the identity of G.
Since the k edges are passed to π in an unknown ordering with unknown labelings, we let V = [k]
and denote the edges as e1 = {u1, v1} , . . . , ek = {uk, vk}. We define an operator next over the domain
{(u, v), (v, u) : {u, v} ∈E} as follows: for {u, v} ∈E, let next(u, v) := (v′, u) where v′ ∈V is the unique
vertex v′ ̸= v such that {u, v′} ∈E. Notice that next is well-defined because all vertices in a cycle have
degree 2. If G = C2, then nextk/2(ui, vi) = (ui, vi) for any i ∈[k].
To set up our encoding of G as a sequence X, we first construct a gadget for each edge ei that will be used
to compute a single next(ui, vi). Under the alphabet Σ = [k] ∪{†, ⋆, _}, we define the nine-token sequence
ei = ⋆ui † vi ui † vi ⋆_.
This gadget ensures that two hops will swap the values of ui and vi. That is
find2
ei◦ui(10) = find1
ei◦ui(6) = 4,
Xfind2
ei◦ui(10) = vi,
find2
ei◦vi(10) = find1
ei◦vi(8) = 2,
Xfind2
ei◦vi(10) = ui.
Likewise, concatenating sequences corresponding to overlapping edges facilitates multiple hops. For example,
if e1 = (1, 2), e2 = (3, 4), e3 = (2, 3), then
find2
e1◦e2◦e3◦2(28) = 22,
Xfind2
e1◦e2◦e3◦2(28) = 3,
find4
e1◦e2◦e3◦2(28) = 13,
Xfind4
e1◦e2◦e3◦2(28) = 4,
find4
e1◦e2◦e3◦3(28) = 2,
Xfind4
e1◦e2◦e3◦3(28) = 1.
Let
E := (e1 ◦e2 ◦· · · ◦ek)k/2 ◦1
be a length Nk := 9k · k
2 + 1 sequence and let X = (_)N−Nk ◦E. We show that hopk(X)N = hopk(E)Nk = 1
if and only if G = C2.
Without loss of generality, let {j, j + 1} = eij ∈E for all j ∈[ k
2 −1]. Let ei0 = {1, v∗}, where v∗= k
2 if
G = C2 and v∗= k if G = C1. Assume without loss of generality that i1 > i0. We argue inductively that for
any j ∈[ k
2]:
31
1. Every two hops simulates a single step of next:
hop2j(E)Nk = nextj(1, v∗)1 =
(
j
if j + 1 < k
2 or G = C1,
1
if j = k
2, G = C2;
2. Every two hops never “jumps” by more than one repetition of all edges gadgets:
find2j
E (Nk) ≥find2j−2
E
(Nk) −9(k −1);
3. The executed gadget corresponds to the correct edge and the gadget is executed correctly:
find2j
E (Nk) ∈{9kj′ + 9ij + ι : j′ ∈N, ι ∈{2, 4}} .
If all three conditions are met, then hopk(X)N = 1 if and only if G = C1 from condition 1.
We first show that the base case holds for j = 1. Since i1 > i0, the second-last time 1 appears in the E is
in the final encoding ei1. By the two-case analysis of the ei1 gadget, we validate that hop2(E)Nk = 2 and
conditions (1) and (3) hold. Since ei1 cannot be the first edge encoding appearing in e1 ◦e2 ◦· · · ◦ek, owing
to it following ei0), condition (2) is satisfied.
Suppose that the inductive hypotheses holds up to j <
k
2. Then, we argue that it holds for j + 1.
Since hop2j(E)Nk = j + 1 (from condition (1)) and find2j
E (Nk) resides at the left-most side of the gadget
for eij (from condition (3)), the two subsequent findE iterations must occur in the gadget eij+1. Because
find2j
E (Nk) ≥9k(k −j) (from condition (2)), all edges appear in the k gadgets to the left of find2j
E (Nk),
and all other edges (including eij+1) must occur before the next occurrence of eij. Thus, the two hops
occur in the eij+1 gadget (within distance 9(k −1)) and results in a properly positioned find2j+2
E
(Nk) with
hop2j+2(E)Nk = nextj+1(1, v∗)1.
Since an MPC protocol can convert G to X using a constant number of layers, and because π′ outputs
T(X)N = 1 if and only if G = C1, we can construct a protocol of π by simulating π′. Because the protocol π
uses s = O(k1−ϵ/2) local memory and q = O(k2/ξ) machines, Conjecture 2.4 implies that the existence of T
requires L = Ω(log k).
E
Proofs from Section 5
E.1
Multi-player pointer chasing communication complexity
We introduce the multi-pass multi-player blackboard communication model studied by Guha and McGregor
(2009) and Assadi and N (2021) to prove lower bounds for multi-pass streaming algorithms. A protocol in this
model specifies how k players, each possessing a portion of a shared input, can jointly compute a function on
the input over the course of R rounds of communication. In each round, all players take turns to broadcast
an s-bit message to all other players. We provide a formal definition of the model as described in Section 6 of
Assadi and N (2021).
Definition E.1. A k-player R-round s-space sequential blackboard communication protocol includes k players
P1, . . . , Pk. On input Z that can be partitioned into (Z1, . . . , Zk), each player Pj is provided with its respective
Zj. In each round, players communicate via a shared blackboard. That is, in round r and in order Pk, . . . , P1,
each player Pj writes a message Πr
j ∈{0, 1}s on the blackboard (which can be viewed by all players) as a
potentially randomized function of input Zj and all information on the blackboard. After the conclusion of R
rounds, the final message ΠR
1 is the output of the protocol.
Assadi and N (2021) proves a lower bound on the round complexity necessary to solve the well-studied
multi-party pointer chasing problem of Nisan and Wigderson (1993). We present the problem as defined by
Assadi and N (2021).
32
Definition E.2. For q, k ∈Z+, let an (q, k)-layered graph G = (V, E) have disjoint vertex layers V1, . . . , Vk+1
with V = V1 ∪· · · ∪Vk+1 and each |Vj| = q and edge layers E1, . . . , Ek with E = E1 ∪· · · ∪Ek and each Ej
being a perfect matching between Vj and Vj+1. The pointer chasing task is provides a (q, k)-layered graph G,
an arbitrary v ∈V1, and an arbitrary equipartition V 1
k+1 and V 2
k+1 of Vk+1 as input and asks whether v is
connected to a vertex in V 1
k+1 or V 2
k+1.
Assadi and N (2021) give the following lower bound.
Proposition E.3 (Proposition 4.12 of Assadi and N, 2021). Consider a k-player R-round s-space sequential
blackboard protocol that solves the (q, k)-pointer chasing task where each player Pj is provided with the
matching Ej and v and V 1
k+1, V 2
k+1 are globally known. Then, the protocol succeeds with probability at least 2
3
only if R ≥k or s = Ω( q
k5 ).
All of the lower bounds in Section 5 are most naturally proved by reducing from hopk, rather than
pointer chasing. So we first prove a lower bound for hopk using the lower bound for pointer chasing from
Proposition E.3.
Proposition E.4. Consider a k-player R-round s-space sequential blackboard protocol that computes
hopk(X)N on any X ∈ΣN for Σ = [2q + 2] with q =
N
2k
where each player Pj is provided with
Xj := (X2(k−j)q+1, . . . , X2(k−j+1)q), except for P1, who is given X1 := (X2(k−1)q+1, . . . , XN). Then, the
protocol succeeds with probability at least 2
3 only if R ≥k or s = Ω( N
k6 ).
Proof. Assuming the existence of a k-player R-round s-space sequential blackboard protocol for hopk(X)N
as described above, we design a protocol for solving (q, k)-pointer chasing with R rounds and s-size messages.
The claimed lower bound will then follow by Proposition E.3.
Consider any pointer chasing input with universally known V1, . . . , Vk+1, v ∈V1, and V 1
k+1 and V 2
k+1, and
each player Pj knowing matching Ej. We recursively define v1, . . . , vk +1 such that v1 = v and (vj, vj+1) ∈Ej,
noting that the output hinges on whether vk+1 ∈V 1
k+1.
Without loss of generality, let v = 1 and
Vj =
(
{1, . . . , q}
if j is odd,
{q + 1, . . . , 2q}
if j is even.
Each player independently determines their substring Xj of a input X to hopk before running the aforemen-
tioned protocol:
• Player P1 encodes X1 by letting XN = s = 1 and for any i ∈1, . . . , 2q, letting
X1
i =
(
i+1
2
∈V1
if i is odd,
i′ ∈V2
if i is even, ( i
2, i′) ∈E1.
This ensures that that every integer in {1, . . . , 2q} appears exactly once in X1
1, . . . , X1
2q, which in turn
guarantees that find1
X(N) = (k −1 + 1)q + 2 and that Xfind1
X(N) = v2 where (1, i′) ∈E1.
• For any j ∈{2, . . . , k −1}, player Pj encodes Ej as Xj as follows.
If j is odd, then for every
i ∈{1, . . . , 2q},
Xj
i =
(
i+1
2
∈Vj
if i is odd,
i′ ∈Vj+1
if i is even, ( i
2, i′) ∈Ej.
Alternatively, if j is even,
Xj
i =
(
q + i+1
2
∈Vj
if i is odd,
i′ ∈Vj+1
if i is even, ( i
2, i′) ∈Ej.
33
Since every odd token corresponds to a vertex in Vj and each subsequent token corresponds to the
vertex it’s connected to by Ej, we can ensure that for every i ∈[2q]:
(X2(k−j+1)+i, Xfind1
X(2(k−j+1)+i)) ∈Ej.
Hence, it follows inductively that Xfindj
X(N) = vj+1.
• Player Pk encodes Xk if k is odd by letting
Xk
i = Xi =
i+1
2
∈Vk
if i is odd,
2q + 1
if i is even, ( i
2, v) ∈Ek, and v ∈V 1
k+1,
2q + 2
if i is even, ( i
2, v) ∈Ek, and v ∈V 2
k+1.
Likewise, if k is even,
Xk
i = Xi =
q + i+1
2
∈Vk
if i is odd,
2q + 1
if i is even, ( i
2, v) ∈Ek, and v ∈V 1
k+1,
2q + 2
if i is even, ( i
2, v) ∈Ek, and v ∈V 2
k+1.
These jointly ensure that
hopk(X)N = Xfindk
X(N) =
(
2q + 1
if vk+1 ∈V 1
k+1,
2q + 2
if vk+1 ∈V 2
k+1.
Therefore, by formatting E1, . . . , Ek appropriately as X, running the protocol for hopk(X)N, and observing
that the final output of player P 1 is 2q + 1 if and only if vk+1 ∈V 1
k+1, there exists a k-player R-round s-space
protocol for pointer chasing. Hence, by Proposition E.3, the protocol for hopk(X)N must use R ≥k rounds
or s = Ω( N
k6 ) space.
E.2
Proofs of Section 5.2
Corollary 5.2. A multi-layer RNN of depth L and width m as above with YN = hopk(X)N satisfies either
L ≥k or m = Ω( N
k6 ).
Proof. Suppose there exists a multi-layer RNN computing output Y with YN,1 = hopk(X)N from input X
with intermediate states Z1, . . . , ZL−1 and hidden states H1, . . . , HL. For any ℓ∈[L] and i ≤i′, note that
Zℓ
i , . . . , Zℓ
i′ can be determined exactly from Hℓ
i−1 and Zℓ−1
i
, . . . , Zℓ−1
i′
. Given this RNN, we provide a multi-
player blackboard communication protocol for solving hopk(X)N under the input model of Proposition E.4.
In round r, we assume inductively that each player Pj knows Zℓ−1,j = (Zℓ−1
2(k−j)q+1, . . . , Zℓ−1
2(k−j+1)q),
except for P1, who knows Zℓ−1,1 = (Zℓ−1
2(k−1)q+1, . . . , Zℓ−1
N
). In descending order, each player Pj computes
Zℓ,j and Hℓ
2(k−j+1)q—writing the latter on the blackboard—from Zℓ−1,j and Hℓ
2(k−j)q,which was written on
the blackboard by the previous player. Thus, player P 1 after round L knows and outputs ZL
N,1 = YN,1 =
hopk(X)N, which provides an L-round protocol m-space protocol.
So the claimed lower bounds on width and depth follow from Proposition E.4.
E.3
Proofs of Section 5.3
Corollary 5.3. Any T ∈KernelFormerN
m,m′,L,H with T(X)N = hopk(X)N satisfies either L ≥k or mm′Hp =
Ω( N
k6 ).
Proof. Under the distribution of input X = (X1, . . . , Xk) to players P1, . . . , Pk stipulated in the statement
of Proposition E.4, we explain how the players can all compute the outcome of a single layer of H-headed
34
kernelized attention in a single round of a blackboard protocol. It is immediate that a depth L network can
be simulated in L rounds.
On input X, consider H kernelized self-attention units with embeddings (Q′
1, K′
1, V1), . . . , (Q′
H, K′
H, VH)
and output MLP ψ. Each player Pj immediately computes its embeddings (Q′
h(Xj), K′
h(Xj), Vh(Xj))h∈[H],
followed by (K′
h(Xj)TVh(Xj)) ∈Rm′×m for each h ∈[H]. Because the object is to compute for each h
ψ(Q′
h(X)K′
h(X)TVh(X)) = ψ(Q′
h(X)
k
X
j=1
K′
h(Xj)TVh(Xj)),
each player writes their (K′
h(Xj)TVh(Xj))h∈[H] using message size s = Θ(mm′Hp). Each can then construct
K′
h(X)TVh(X)) by reading the board, and use it to compute its respective outputs without requiring
supplemental communication.
Hence, T (and thus hopk(X)N) can be simulated using an L-round blackboard protocol with message
size s = Θ(mm′Hp), and the corollary follows from Proposition E.4.
Corollary 5.4. Any T ∈Λw,g-AttnN
m,L,H with T(X)N = hopk(X)N satisfies either L ≥k or (w+ N
gk)mHp =
Ω( N
k6 ).
Proof. As in the proof of Corollary 5.3, we explain how each player can compute their respective outputs of a
single unit of self-attention masked by Λw,g.
To compute the output corresponding to Xi, note that it is necessary to only know the embeddings
corresponding to Xi−w, Xi−w+1, . . . , Xi+w and Xg, X2g, . . . , X⌊N/g⌋g. Thus, player Xj can compute the
outputs of all of their inputs Xj = (X2(k−j)q+1, . . . , X2(k−j+1)q) given access to
X2(k−j)q+1−w, . . . , X2(k−j)q, X2(k−j+1)q+1, . . . , X2(k−j+1)q+w,
as well as Xg, X2g, . . . , X⌊N/g⌋g.
Therefore, the protocol can be simulated if each player Xj writes inputs
X2(k−j)q+1, . . . , X2(k−j)q+w, X2(k−j+1)q−w+1, . . . , X2(k−j+1)q ∈Rm,
in addition to all Xi ∈Xj such that i ≡0 (mod g). This can be accomplished by a protocol where each
player writes s = O((w + N
gk)mp) bits of information on the blackboard.
By repeating this protocol in parallel for every head and sequentially for every layer, T and hopk(X)N
can be simulated, and hence the claim follows from Proposition E.4.
E.4
Proofs of Section 5.4
Corollary 5.6. Any T ∈MaskTransformerN+NCoT
m,1,H
that computes hopk(X)N with NCoT tokens of chain-of-
thought requires either NCoT ≥k or mHp = Ω( N
k6 ).
Proof. We reduce to Proposition E.4. Consider some input X ∈RN partitioned into X1, . . . , Xj as specified
by the proof of Proposition E.4 with chain-of-thought XCoT and hopk(X)N determined by some masked
transformer T.7 Suppose T has embeddings (Qh, Kh, Vh)h∈[H] and output MLP ψ. We provide an (NCoT +1)-
round blackboard protocol to compute hopk(X)N from X.
Suppose in the rth round of the protocol, all players know XCoT,1, . . . , XCoT,r−1 and aim to compute
T(X ◦XCoT)N+r−1 =
(
XCoT,r
if r ≤NCoT
hopk(X)N
if r = NCoT + 1
= ψN+r−1
XN+r−1 +
H
X
h=1
PN+r−1
i=1
exp(Qh
N+r−1(XN+r−1)TKh
i (Xi)T)V h
i (Xi)
PN+r−1
i=1
exp(Qh
N+r−1(XN+r−1)TKh
i (Xi))
!
.
7We abuse notation to index XN+i = XCoT,i and let Xi ∈Xj be true if i ∈{2(k −j)q + 1, . . . , w(k −j + 1)q}.
35
If we let
Sr,h,j =
X
Xi∈Xj
exp(Qh
N+r−1(XN+r−1)TKh
i (Xi)T)V h
i (Xi) ∈Rm,
Sr,h,CoT =
N+r−1
X
i=N+1
exp(Qh
N+r−1(XN+r−1)TKh
i (Xi)T)V h
i (Xi) ∈Rm,
Zr,h,j =
X
Xi∈Xj
exp(Qh
N+r−1(XN+r−1)TKh
i (Xi)T) ∈R,
Zr,h,CoT =
N+r−1
X
i=N+1
exp(Qh
N+r−1(XN+r−1)TKh
i (Xi)T) ∈R,
then we observe that
T(X ◦XCoT)N+r−1 = ψN+r−1
XN+r−1 +
H
X
h=1
Pk
j=1 Sr,h,j + Sr,h,CoT
Pk
j=1 Zr,h,j + Zr,h,CoT
!
.
Each player Pk computes (Sr,h,j, Zr,h,j)h∈[H] and writes them on the blackboard with O(mHp)-bit messages.
Since Sr,h,CoT and Zr,h,CoT are known by all players, every player can individually T(X ◦XCoT)N+r−1.
By induction, all players know hopk(X)N after NCoT + 1 rounds. The claim now follows from Proposi-
tion E.4.
F
Proofs of low-level attention constructions
F.1
Hardmax simulation proof of Appendix A.1
Lemma A.2. Let f ∈AttnN
m be a self-attention unit with precision p = Θ(log N) and embedding functions
Q, K, V such that for some fixed 1 ≥ξ = N −O(1) and every X ∈RN×m and i ∈[N]:
A(X)i,i′ ≤max
i′′ A(X)i,i′′ −ξ, ∀i′ ̸∈Imax(A(X)i),
where A(X) = Q(X)K(X)T. Then there exists a self-attention unit f ′ ∈AttnN
m with a valid p′-bit implemen-
tation with p′ = O(p) satisfying
f ′(X) = hardmax(A(X))V (X).
Proof. For some p′ = Θ(p + log 1
ξ ) and c = Θ( p′+ζ
ξ
· log N) where ζ is as in Appendix A.1), let f ′ have query
embedding Q′(X) = cQ(X) and identical key K and value V embeddings as f. Therefore, by construction,
these embeddings can be written with precision p′ = O(ln(c) + p) = O(log 1
ξ + log log N + p) = O(p).
Let ˆf ′ be a valid p′-bit implementation of f ′, meaning that the two ∥ˆf ′ −f ′∥∞= O(1/2p+1) (thus ˆf ′
rounds f ′ to p′ bits of precision), and fix some X. We first show that the softmax matrix is sufficiently close
to that of the hardmax and is also a valid p′-bit implementation of the hardmax. Without loss of generality,
let 1 ∈Imax(A(X)i). First, note that
X
i′̸∈Imax(A(X)i)
exp(cA(X)i,i′) ≤
N
exp(cξ) exp(cA(X)i,1) =
1
N O(p′+ζ) exp(cA(X)i,1).
Then,
|softmax(cA(X))i,1 −hardmax(A(X))i,1| =
1
|Imax(A(X)i)| −
exp(cA(X)i,1)
PN
i′=1 exp(cA(X)i,i′)
≤
P
i′̸∈Imax(A(X)i) exp(cA(X)i,i′)
|Imax(A(X)i)| exp(cA(X)i,1)
=
1
N Ω(p′+ζ) .
36
Likewise, for any i′′ ̸∈Imax(A(X)i):
|softmax(cA(X))i,i′′ −hardmax(A(X))i,i′′| ≤
exp(cA(X)i,i′′)
PN
i′=1 exp(cA(X)i,i′)
=
1
N Ω(p′+ζ) .
Therefore,
∥softmax(cA(X))i −hardmax(cA(X))i∥2 ≤
√
N·max
i′′ |softmax(cA(X))i,i′′ −hardmax(cA(X))i,i′′| =
1
N Ω(p′+ζ) .
We conclude that the approximation is sufficiently close, meaning it is O(1/2p′), whereby it is exact after
rounding:
ˆf ′(X) −hardmax(Q(X)K(X)T)V (X)
∞≤
f ′(X) −hardmax(Q(X)K(X)T)V (X)
∞+
ˆf ′(X) −f ′(X)
∞
≤max
i,j
softmax(cA(X))T
i V (X)·,j −hardmax(A(X))T
i V (X)·,j
+ O
1
2p′
≤max
i,j
softmax(cA(X))T
i −hardmax(A(X))T
i
2 ∥V (X)·,j∥2 + O
1
2p′
≤
1
N Ω(p′+ζ) ·
√
N · N ζ + O
1
2p′
= O
1
2p′
.
Therefore, ˆf ′ is a valid p′-bit implementation of hardmax(Q(X)K(X)T)V (X).
F.2
Constructions for Appendix B.1
Proposition B.1. For any b ≤N and d, there exists a self-attention unit sparsePropagateQ,d ∈AttnN
m,p for
m = d+O(Q log N) and p = O(log N), which, given any input X with Xi = (zi, Si,⃗0) ∈Rd×
[N]
≤Q
×{0}m−Q−d
such that bi = |{Sj ∋i : j ∈[N]}| ≤Q for all i, has output sparsePropagateQ,d(X) satisfying
sparsePropagateQ,d(X)i = 1
bi
X
Sj∋i
zj.
Proof. Following the proof of Theorem 2 of Sanford et al. (2023), there exist p-bit precision vectors u1, . . . , uN ∈
{±1/√m}m and wS with wS ≤2√Q for all S ∈
N
≤Q
such that
uT
i wS = 1, for all i ∈S
uT
i wS ≤1
2, for all i ̸∈S.
We then design the embeddings of sparsePropagateQ,d with
Q(X)i = (ui, 1),
K(X)i =
(
(wSi, 0)
if i > 0,
(⃗0, 3
4)
if i = 0,
V (X)i =
(
zi
if i > 0,
⃗0
if i = 0.
As a result,
Q(X)T
i K(X)i′ = 1
if i ∈Si′, i′ > 0,
Q(X)T
i K(X)i′ ≤1
2
if i ̸∈Si′, i′ > 0,
Q(X)T
i K(X)0 = 3
4.
37
Hence, the largest inner products for query i correspond to i′ for all Si′ ∋i if any exist, and 0 otherwise.
There exists a margin of at least 1
4 between the largest inner product in each row and all others. By applying
Lemma A.2, we conclude that there exists a self attention unit f ′ with embedding dimension p = Θ(log N)
that computes
f ′(X) = hardmax(Q(X)K(X)T)V (X) = sparsePropagate(X).
F.3
Constructions for Appendix B.2
Lemma B.4. For any MPC protocol π with local memory s and q machines with nin-word inputs, there
exists a transformer init ∈Transformernin,max(nin,q)
s,1,1,din,dout
with din = 1 and dout = s, which, given Input ∈Zn
2p, has
output satisfying init(Input) = MachineIn(1).
Proof. Let M = max(nin, q) and Q, K, V : ZM
2p →RM×s be the query, key, and value embeddings of the
attention unit f in init, and let ψ : RM×s →Zs
2p × [N] be its output MLP. Let qin =
nin
s
denote the number
of machines used to store the inputs.
Let Desti′ =
l
i′
s
m
∈[qin] denote the machine that stores the input token index i′ ∈[nin] in the MPC
protocol, and let
Rcvdi = {(s −1)i + 1, . . . , min(si, nin)}
denote the set of all input tokens indices belonging to MachineIn(1)
i
for machine i ∈[qin].
For each machine i ∈[qin], we define the query embedding as
Q(Input)i =
cos
2πi
M
, sin
2πi
M
, . . . , cos
2πi
M
, sin
2πi
M
.
Likewise, for each token index i′ ∈[nin], the key and value vectors are
K(Input)i′,(2ι−1,2ι) =
(
cos
2π·Desti′
M
, sin
2π·Desti′
M
if i′ ≤nin, i′ ≡ι
(mod s),
(0, 0)
otherwise,
V (Input)i′,(2ι−1,2ι) =
(
(Inputi′, i′)
if i′ ≤nin, i′ ≡ι
(mod s),
(0, i′)
otherwise.
These definitions guarantee that large inner products only occur between machine queries Q(Input)i and
tokens keys K(Input)i′ when Inputi′ is allocated to MachineIn(1)
i . That is,
Q(Input)T
i K(Input)i′ = 1,
if i′ ∈Rcvdi
Q(Input)T
i K(Input)i′ ≤1 −Ω
1
M 2
,
otherwise.
By applying Lemma A.2 with ξ = Ω( 1
N2 ), there exists some self-attention unit f ′ such that
f ′(Input)i = hardmax(Q(Input)K(Input)T) = (Inputi′, i′)i′∈Rcvdi
|Rcvdi|
.
A proper choice of ψ and an invocation of the definition of MachineIn(1) ensures that init(Input)i =
ψ(f(Input))i = MachineIn(1)
i .
Lemma B.6. For any R-round MPC protocol π with local memory s and q machines with nout-word output,
there exists a transformer final ∈Transformerq,max(nout,q)
s,1,1,din,dout
for din = s and dout = 1, which, given input
X = MachineIn(R), has output final(X) with final(X)i,1 = Outputi ∈Z2p.
38
Proof. This argument inverts that of Lemma B.4, after applying the LocalR to transform MachineIn(R) to
MachineOut(R). Let Q, K, V : ZM
2p →RM×s be the query, key, and value embeddings of the only attention
unit f in final, and let ψ : RM×s →Zs
2p × [N] be its output MLP. Let qout =
nout
s
denote the number of
machines storing relevant information for the output of the MPC protocol.
For each machine i′ ∈[qout], let
Senti′ = {(s −1)i′ + 1, . . . , min(si′, nout)}
denote the set of all token indices receiving its output. Likewise, for each token index i ∈[nout], let Srci = ⌈i/s⌉
be the machine containing its relevant token. We define Q = Q′ ◦LocalR, K = K′ ◦LocalR, V = V ′ ◦LocalR
as follows.
Q′(MachineOut(R))i,(2ι−1,2ι) =
(
cos
2π⌊Srci⌋
M
, sin
2π⌊Srci⌋
M
if i ≤nout, i ≡ι
(mod s)
(0, 0)
otherwise.
K′(MachineOut(R))i′ =
cos
2πi′
M
, sin
2πi′
M
, . . . , cos
2πi′
M
, sin
2πi′
M
.
V ′(MachineOut(R))i′ = MsgOut(R)
i′ .
Applying Lemma A.2 as before yields
f(MachineIn(R))i =
(
MachineOut(R)
i′
if i ∈Senti′,
0
otherwise.
A properly chosen ψ ensures that final(MachineIn(R))i = ψ(f(MachineIn(R)))i = Outputi.
F.4
Constructions for Appendix D.1
Lemma D.1. For some m ≥d + 2, τ : [N] × Rm →[N], and ρ : Rm →Rd, there exists an attention head
lookUpτ,ρ ∈MaskAttnN
m with precision p = O(log N) and m ≥d+2 satisfying lookUpτ,ρ(X)i,:d = ρ(Xτ(i,Xi)).
Proof. We let V (Xi) = (ρ(Xi),⃗0) and define sinusoidal embeddings Q and K with
Q(X)i =
cos
2πτ(i, Xi)
N
, sin
2πτ(i, Xi)
N
,⃗0
,
K(X)i =
cos
2πi
N
, sin
2πi)
N
,⃗0
.
Note that
Q(X)T
i K(X)i′ = 1,
if τ(i, Xi) = i′,
Q(X)T
i K(X)i′ ≤cos
2π
N
= 1 −Ω
1
N 2
,
otherwise.
By applying Lemma A.2 with ξ = Ω( 1
N2 ), we conclude that a satisfactory self-attention unit exists.
Lemma D.2. For any finite alphabet Σ, m ≥d + 2, µ1, µ2 : Rm →Σ, and ρ : Rm →Rd, there exists an
attention head lastOccurrenceµ,ρ ∈MaskAttnN
m with precision p = O(log(N |Σ|)) such that,
lastOccurrence(X)i,:d =
(
ρ(⃗0)
if ∀i′ < i : µ1(Xi′) ̸= µ2(Xi),
ρ(Xi′)
if i′ = max {i′ < i : µ1(Xi′) = µ2(Xi)} .
39
Proof. Let N ′ = N|Σ|. We define token embeddings as follows, including start token “dummy embeddings”
as discussed in Appendix A.1.
Q(X)i =
cos
2π(Nµ2(Xi) + i)
N|Σ|
, sin
2π(Nµ2(Xi) + i)
N|Σ|
, 1,⃗0
,
K(X)i =
cos
2π(Nµ1(Xi) + i)
N|Σ|
, sin
2π(Nµ1(Xi) + i)
N|Σ|
, 0,⃗0
,
K(X)0 =
0, 0, cos
2π(N −1
2)
N|Σ|
,⃗0
,
V (X)i = (ρ(Xi),⃗0),
V (X)0 = ⃗0.
Taken together, these embeddings provide the following characterization of the inner products (with causal
masking matrix Γ):
Q(X)T
0 K(X)i′ + Γi,i′ = cos
2π(i −i′)
N|Σ|
if i ≥i′ > 0, µ1(Xi′) = µ2(Xi),
Q(X)T
i K(X)i′ + Γi,i′ ≤cos
2π
N
if i ≥i′ > 0, µ1(Xi′) ̸= µ2(Xi),
Q(X)T
i K(X)i′ + Γi,i′ = −∞
if i < i′,
Q(X)T
i K(X)i + Γi,0 = cos
2π(N −1
2)
N|Σ|
.
As a result, the largest inner product Q(X)T
i K(X)i′ for some i is the largest i′ with µ1(Xi′) = µ2(Xi) if one
exists and i′ = 0 otherwise. Furthermore, there exists a margin of Ω(
1
N2|Σ|2 ) between this inner product and
all others. We conclude by applying Lemma A.2.
G
Further empirical analysis of k-hop induction heads
This appendix presents in-depth explanations of the empirical results of Section 4.2, along with further
experiments. Taken together, these results suggest that the relationship between the number of hops k and
the depth L of transformers trained on the task is well-characterized by the representational thresholds of
Theorem 4.2 and Corollary 4.3; that the construction described in the proof of Theorem 4.2 is attainable by
trained models; and deep models likely exhibit an inductive bias that favors compositional learning rules in
the finite sample regime.
We define our experimental methodology precisely in Appendix G.1 and provide supporting evidence for
our claims in the subsequent sections.
Exponential powers of depth.
Our principal empirical claim is that incrementing the depth L of a
transformer exponentially increases the model’s capabilities to learn k-hop induction heads tasks. We explore
this claim primarily in Appendix G.2, where we compare this empirical claim with the relevant theoretical
results (Theorem 4.2 and Corollary 4.3), which suggest a similar dependence. We further study the impacts
of increasing the embedding dimension m of the transformer in Appendix G.3 and find that doubling the
width is roughly equivalent in performance to incrementing the depth by one.
Empirical Claim G.1. A transformer T ∈MaskTransformerN
m,L,H trained with Adam to solve hopk has
small token-wise classification error if L log(m) = Ω(log k) and large error if L log m = O(log k).
Mechanistic alignment with theoretical construction.
We further demonstrate the empirical salience
of our theoretical construction by conducting a study of the interpretability of learned transformers in
Appendix G.4. This investigation reveals that the attention matrices of sufficiently deep transformers exhibit
40
an implementation of a circuit that relies on the same “doubling” principle of the construction in the proof of
Theorem 4.2. The resulting circuit is comprised of the same intermediate products that are used in that hopk
construction.
Empirical Claim G.2. The outputs of individual attention matrices of a transformer T ∈MaskTransformerN
m,L,H
trained with Adam to solve hopk with L = Ω(log k) and evaluated on input X ∈ΣN (i) correspond to the
findj
X intermediate products of the Theorem 4.2 construction and (ii) demonstrate a “doubling” phenomenon
where the each head layer ℓcorresponds to findj
X for some j = O(2ℓ).
Beneficial inductive biases of depth.
While most of our experiments belong to the “infinite-sample”
regime where new samples are randomly generated on each training step, we also evaluate our models in two
finite-sample regimes in Appendix G.5. We find that a small number of samples is sufficient to approach the
performance of the infinite-sample regime. When the amount of training data is small, we find that deeper
models perform better than shallower models, possibly due to an inductive bias that favors compositional
hypotheses.
Empirical Claim G.3. hopk can be learned in a sample-efficient manner by transformers T ∈MaskTransformerN
m,L,H
trained with Adam with L = Ω(log k). If T overfits to hopk tasks for some k, then increasing the depth L
while holding k fixed leads superior performance.
The experiments detailed here were conducted under limited computational resources. The authors are
interested in future work that would evaluate whether these scaling rules persist on larger architectures and
more complex tasks.
G.1
Experimental details
Task details.
We study a multi-task variant of k-hop induction heads that predicts hopk(X) = (0, hopk(X′))
from input X = (k, X′) for k ∈{0, 1, . . . , kmax}8 and X′ ∈ΣN−1. We refer to this task as multi-hop and
provide the task hyper-parameters in Table 1.
Hyperparameter
Value
Context length N
100
Alphabet size |Σ|
4
Max hops kmax
16
Table 1: Multi-hop task hyper-parameters
We define the distribution Dmulti−hop over labeled samples for the multi-hop task and DX over input
sequences X ∈ΣN−1. We draw a labeled sample (X, hopk(X)) ∼Dmulti−hop by independently sampling
k ∼Unif({0, 1, . . . , kmax}) and X′ ∼DX . Input sequences X′ ∼DX are drawn uniformly from inputs with no
repeating elements. That is, we sample X′
1 ∼Unif(Σ) and each X′
j+1 ∼Unif(Σ \
X′
j
). For each k ∈[kmax],
let Dhopk denote the conditional distribution ((k′, X′), (0, hopk′(X′))) ∼Dmulti−hop | (k = k′). Also, let
dom(hopk) = {(k, X′) : Pr [X′ ∼DX ] > 0}.
For Σ := Σ ∪[kmax], we define the n-sample empirical token-wise classification error of a transformer
T : Σ
N →Σ
N on a task hopk as
errn
k(T) = 1
n
n
X
ι=1
1
| {i : hopk(Xι)i ̸=⊥} |
N
X
i=1
1 {T(Xι)i ̸= hopk(Xι)i ̸=⊥} ,
for iid samples (X1, hopk(X1)), . . . , (Xn, hopk(Xn)) ∼Dhopk. We ignore null ⊥outputs of hopk when no
k-hop induction head exists in order to avoid inadvertently over-estimating the performance of transformers
on large k tasks, which have a large fraction of null outputs.
8The task hop0 is simply the identity mapping: hop0(X′) = X′.
41
Training details.
We trained a variety of causally-masked GPT-2 transformers (Radford et al., 2019) from
HuggingFace to solve the multi-hop task. The model has an absolute positional encoding.
The transformers are trained with Adam (Kingma and Ba, 2014) on the cross-entropy loss. In the
infinite-sample regime, we draw 32 new iid samples from Dmulti−hop on each training step. Otherwise, ntrain
samples are drawn before training commences and all samples are rotated through batches, before repeating.
We use the hyper-parameters in Table 2 to train all of the models identified in Table 3.
Hyperparameter
Value
Embedding dimension m
{128, 256}
Depth L
{2, 3, 4, 5, 6}
Number of heads H
{4, 8}
Vocabulary size
30
Activation function
GeLU
Layer norm ϵ
10−5
Training samples ntrain
103, 3 · 103, ∞
Learning rate
10−4
Training steps
105
Batch size
32
Table 2: Model and training hyper-parameters
Identifier
Heads H
Embedding dimension m
Depth L
Training samples ntrain
Total parameters
T ∞
4,2
4
128
2
∞
413,440
T ∞
4,3
4
128
3
∞
611,712
T ∞
4,4
4
128
4
∞
809,984
T ∞
4,5
4
128
5
∞
1,008,256
T ∞
4,6
4
128
6
∞
1,206,528
T ∞
8,2
8
256
2
∞
1,613,312
T ∞
8,3
8
256
3
∞
2,403,072
T ∞
8,4
8
256
4
∞
3,192,832
T ∞
8,5
8
256
5
∞
3,982,592
T ∞
8,6
8
256
6
∞
4,772,352
T 3000
4,2
4
128
2
3000
413,440
T 3000
4,3
4
128
3
3000
611,712
T 3000
4,4
4
128
4
3000
809,984
T 3000
4,5
4
128
5
3000
1,008,256
T 3000
4,6
4
128
6
3000
1,206,528
T 1000
4,2
4
128
2
1000
413,440
T 1000
4,3
4
128
3
1000
611,712
T 1000
4,4
4
128
4
1000
809,984
T 1000
4,5
4
128
5
1000
1,008,256
T 1000
4,6
4
128
6
1000
1,206,528
Table 3: Hyper-parameters of all MaskTransformerN
m,L,H trained for the empirical analysis.
Computational resources.
All experiments were run on a 2021 Macbook Pro with an M1 chip.
42
G.2
Exponential increases in k-hop capacity with depth (Empirical Claim G.1;
Figures 6 to 8)
1
2
4
8
16
k
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Error: errk
n(T)
Evaluation of L-depth, 4-headed, infinite-sample intransformers on hopk
L = 2
L = 3
L = 4
L = 5
L = 6
Figure 6: Zoomed in version of Figure 2. Evaluation of transformers errn
k(T ∞
4,L) with depths L ∈{2, 3, 4, 5, 6},
heads H = 4, and embedding dimension m = 128 trained on the multi-hop task. This figure plots errn
k(T ∞
4,L)
on n = 100 samples as a function of k for each choice of L.
We visualize the relationship between the depth L of a transformer and the largest k such that errn
k(T)
is small in Figure 6, Figure 7, and Figure 8. We exhibit the relationship in its simplest form by considering
transformers with heads H = 4, embedding dimension m = 128, and new training samples on every epoch.
The figures provide alternate views of errn
k(T ∞
4,L) for each L ∈{2, 3, 4, 5, 6} with n = 100 samples for each
k ∈[kmax].
Together, these plots illustrate a sharp phase transition when D = ⌊log2 k⌋+ 2, which identically matches
the depth scaling in Theorem 4.2. Increasing the depth of a transformer by one approximately doubles
the number of values k ∈[kmax] with bounded error. For instance, following the theoretical and empirical
intuition of Bietti et al. (2023), the depth L = 2 transformer T ∞
4,2 succeeds in solving the standard induction
heads task, but attains at least 10% error on all other tasks. Likewise, a depth L = 3 model has error bounded
by 1% for k ∈{1, 2}, which increases rapidly for larger values of k.
This doubling phenomenon suggests that simple compositional tasks with a larger number of compositions
than the depth of the model are easily learnable if the model can employ a doubling trick, similar to the one
used in the proof of Theorem 4.2. This relationship between compositionality and depth reflects the results
of Zhang et al. (2023), where the learnable task complexity also scales super-linearly in depth.
Given the lower bounds of Corollary 4.3, one may ask why models with depth L < ⌊log2 k⌋achieve
non-trivial success on hopk tasks that cannot be represented in a compositional manner. There are several
relevant explanations:
1. In these experiments, the embedding dimension m = 128 is actually larger than the context N = 100,
which may enable the model to memorize more of its preceding samples and offload logical work to the
MLP, rather than executing a pointer-doubling strategy. While practical models regularly have the
opposite (and our theoretical results are oriented around that parametric scaling), we used a larger
43
2
3
4
5
6
Depth L
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Error: errk
n(T)
Evaluation of L-Depth, 4-headed, infinite-sample transformers on hopk
k = 1
k = 2
k = 3
k = 4
k = 5
k = 6
k = 7
k = 8
k = 9
k = 10
k = 11
k = 12
k = 13
k = 14
k = 15
k = 16
Figure 7: Alternate view of Figure 6 including errn
k(T ∞
4,L) plotted as a function of L for each k.
m than is necessary for representational purpose to improve the optimization landscape and speed
convergence.
2. This is made further plausible by the small alphabet size |Σ| and randomly drawn sequences X′, which
place effective bounds on how much look-back from each token i is necessary to compute hopk(X)i.
Nonetheless, these results provide strong support that models are substantially easier to train to low
classification error in the regime where the depth is sufficient to implement a pointer-doubling construction.
In the following subsection, we further investigate this phenomenon by examining the intermediate attention
matrices produced by trained models.
44
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
k
2
3
4
5
6
Depth L
0.01
0.14
0.35
0.42
0.54
0.59
0.62
0.66
0.69
0.69
0.71
0.72
0.73
0.73
0.74
0.74
0.003 0.008 0.048
0.15
0.28
0.36
0.42
0.46
0.52
0.57
0.6
0.64
0.67
0.67
0.69
0.71
0.001 0.003 0.005 0.015 0.034 0.062
0.11
0.22
0.28
0.36
0.4
0.47
0.53
0.53
0.58
0.61
0.001 0.001 0.001 0.003 0.012 0.015 0.017 0.037 0.058
0.1
0.15
0.22
0.28
0.34
0.38
0.41
0
0
0.001 0.002 0.003 0.003 0.007 0.006 0.007
0.01
0.011 0.012 0.021 0.027 0.035 0.062
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Figure 8: Alternate views of Figure 6 including errn
k(T ∞
4,L) as a table with one cell for each (L, k) pair.
G.3
Width variation (Empirical Claim G.1; Figure 9)
While the primary focus of these empirical results and the paper as a whole is on the role of depth in the
ability of transformer to learn parallelizable and compositional tasks, we also aim to understand the interplay
of depth and width in learning the multi-hop task. Here, we contrast the previous transformers T ∞
4,L with
models T ∞
8,L that have more heads (H = 8) and larger embedding dimensions (m = 256). We plot the
classification errors of all 10 architectures over 16 hopk sub-tasks in Figure 9.
Here, we observe a rough correspondence in performance between the transformers T ∞
H,L and T2H,L−1 and
the same doubling phenomenon as is evident models with H = 4 heads. That is, while increasing the width
improves the classification error of learned models, it does so in a far less parameter-efficient manner than
incrementing the depth. As mentioned before, the relative success of wide and shallow transformers is likely
contingent on the relatively short context length N and alphabet size |Σ|. However, these results still suggest
an important role for wider models to play beyond representational capabilities of transformers.
45
1
2
4
8
16
k
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Error: errk
n(T)
Evaluation of L-depth, H-headed, infinite-sample intransformers on hopk
L = 2, H = 4
L = 2, H = 8
L = 3, H = 4
L = 3, H = 8
L = 4, H = 4
L = 4, H = 8
L = 5, H = 4
L = 5, H = 8
L = 6, H = 4
L = 6, H = 8
Figure 9: Comparison between the errors errn
k(T ∞
H,L) of transformers with embedding dimension and heads
(m, H) = (4, 128) (dashed line, same plots as Figure 6) and (m, H) = (8, 256) (solid line) trained on the
multi-hop task, evaluated on n = 100 samples per hopk task.
46
G.4
Mechanistic alignment with theoretical construction (Empirical Claim G.2,
Figures 10 to 15)
We use standard attention-based interpretability techniques to better understand what particular logical
circuits are implemented by transformers trained to solve the multi-hop task. By qualitatively inspecting
the attention matrices produced by trained models and by measuring the alignment between those inner
products and partial solutions findj of hopk, we uncover a striking correspondence between the behaviors
of the trained models and the transformer construction designed in the proof of Theorem 4.2. We further
observe that trained transformers with high accuracy have “decisive” self-attention units with particularly
strong correlations to some findj intermediate, while poorly performing models have less predictable attention
activations.
For a fixed trained model T ∈TransformerN
m,L,H, we let Aℓ,h[T](X) represent the output of the hth self-self
attention matrix in the ℓth layer for h ∈[H] and ℓ∈[L], evaluated at some input X ∈dom(hopk). That is,
we let
Aℓ,h[T](X) = softmax
Qℓ,h(Xℓ−1)Kℓ,h(Xℓ−1)T + Γ
∈RN×N,
where Xℓ−1 is the intermediate state representing the output of layer ℓ−1 of T on input X and Γ is the
causal masking matrix. Each row i in the matrix represents the coefficients of the convex combination of
value vectors affiliated with each query, which can be used as a signifier of which embeddings i receives
information from.
Visualization of findj alignment for hop16 and depth L = 6 (Figure 10).
The outputs of self-attention
matrices are often highly structured matrices that reveal which relationships between tokens are encoded and
how information is shared within the model (Li and McClelland, 2022; Clark et al., 2019; Rogers et al., 2021).
We plot several self-attention matrices associated with a depth L = 6, heads H = 4 transformer trained in
the infinite-sample regime and evaluated on a single sample X ∈dom(hop16) in Figure 10.
By looking at the six self-attention matrices, one can infer that all heads are “decisive” and obtain nearly
all of their relevant information from a single value embedding, rather than averages of a large number
of embeddings. The top-left self-attention matrix, which belongs to the first self-attention head, clearly
associates elements with their predecessors, which is identical the to the function of our lookUp attention
head in the first layer of the hopk construction of Theorem 4.2.
While the roles of the other heads are not immediately obvious, they can be understood by overlaying
colored matrices with non-zero cells at (i, findj
X(i)) for some j ≤k. For instance, the top-right attention matrix
in layer ℓ= 2 corresponds almost exactly with find1
X (as suggested by the second-layer of our construction),
and the others are closely associated with find1
X, find2
X, find3
X, and find8
X for layers ℓ= 3, 4, 5, 6 respectively.
This is a remarkably close correspondence to our construction, which includes a self-attention matrix in the
ℓth layer whose activations correspond to find2ℓ−2
X
.
While not conclusive, this experiment suggests a strong alignment between the behaviors of this particular
transformer and our theoretical construction. This suggests a high likelihood that the transformer successfully
learns to solve hop16 by employing a pointer-doubling primitive. However, these results apply to only a single
model, a single task, and a single input; in the subsequent section, we generalize this interpretability analysis.
Alignment between attention heads and findj for a single hopk sub-task (Figures 11 to 13).
To
broaden and quantify the analysis of the previous section, we measure the extent to which each self-attention
head mimics the functionality of findj, which are partial computations of hopk that are employed in the proof
of Theorem 4.2. We use cell-wise matrix inner products to quantify the strength of correlation between a
self-attention matrix and a fixed function potentially relevant to interpretability.
For two matrices A, B ∈RN×N, let
⟨A, B⟩= ∥A ⊙B∥2
F
∥A∥F ∥B∥F
be their normalized element-wise inner-product, where ∥·∥F is the Frobenius norm and ⊙denotes element-wise
47
= 1, h = 1, find0
X highlighted
= 2, h = 4, find1
X highlighted
= 3, h = 2, find1
X highlighted
= 4, h = 2, find2
X highlighted
= 5, h = 3, find3
X highlighted
= 6, h = 1, find8
X highlighted
Self-attention matrix A , h[T4, 6](X), X
dom(hop16)
Figure 10: The outputs of several internal self-attention matrices Aℓ,h[T ∞
4,6](X) ∈R100×100 of a trained
multi-task transformer of depth D = 6 evaluated on a single sample X ∼Dhop16 are plotted in grayscale. In
each cell, the matrix with non-zero entries (findj
X(i), i)i∈[N] for some j is included in transparent color to
visualize the function of each self-attention unit.
48
multiplication. For some function g : [N] →{0} ∪[N], we let ⟨g, B⟩:= ⟨Ag, B⟩, where
Ag
i,j =
(
1
if g(j) = i,
0
otherwise.
We use this notation to analyze experimentally how closely the self-attention matrices Aℓ,h encode the
intermediate products of the proof of Theorem 4.2, findj
X. For n iid samples X1, . . . , Xn ∈∼Dhopk, let
Aℓ,h, findj
n,k := 1
n
n
X
ι=1
D
findj
Xι, Aℓ,h(Xι)
E
.
Due to the non-negativity of Aℓ,h and findj,
Aℓ,h, findj
n,k ∈[0, 1], and
Aℓ,h, findj
n,k = 1 only if ∀ι ∈[n]:
Aℓ,h(Xι)i,i′ = 1
⇐⇒
findj
Xι(i) = i′.
These inner products make it possible to visualize the strength of correlations of all heads in a particular
model T ∈MaskTransformerN
m,L,H with all target functions findj on a collection of random samples drawn
from some Dhopk. Figure 11 visualizes the functionality of all attention units in the 4-layer, 4-head transformer
T ∞
4,4 when evaluated on the sub-task hop4. The figure gives several clues about how hop4 is successfully
computed by the trained model: the second layer and third layer both utilize find1 to determined find2 jointly
by the end of the third layer. The fourth layer uses the ability to create a stable find2 construction to obtain
find4 and hence hop4.
This plot also indicates the relative stability of this circuit interpretation of the procedure: a large number
of heads are very strongly correlated with find1 or find2 across the 10 samples, which indicates they are likely
utilized consistently to compute those intermediates regardless of input.
Figure 12 is a similar plot for the transformer T ∞
4,6 with depth L = 6, evaluated on the task hop16. The
functionalities of the heads visualized in Figure 10 can be observed in the corresponding inner products. The
collection of all inner products presents further evidence that the pointer-doubling phenomenon occurs in the
trained models, due to the increase in compositions present in the largest inner products of deeper attention
units.
While Figures 11 and 12 showcase the decisive alignment between self-attention heads and particular
partial computations findj in successfully trained models, Figure 13 demonstrates the loss of that decisiveness
in poorly performing transformers. There, we visualize the alignments of the trained depth-4 transformer
T ∞
4,4 evaluated on hop16, in which it attains a 61% token error. While a self-attention units in the second
layer coincides with find1, no strong correlations emerge deeper in the model. Unlike the other figures, the
deeper self-attention units are “indecisive,” lacking any large inner products and failing in particular to
correlate with any highly compositional targets. This provides a visual explanation of the transformer’s
failure, since it lacked the effective representational capacity needed to learn a circuit with consistent and
highly-compositional outputs.9
Alignment between attention heads and findj for all hopk sub-tasks (Figures 14 and 15).
For
an even more global lens on the mechanistic interpretability of these trained models, we visualize how the
maximum inner products of each self-attention unit change for a fixed transformer for different sub-tasks
hopk. Figures 14 and 15 do so for the depth-4 and depth-6 networks respectively. The hue of each cell (and its
numerical label) corresponds to the j∗with the most correlated inner product with corresponding attention
unit Aℓ,h in samples from dom(hopk), and the opacity corresponds to the magnitude of that inner product.
The takeaways of the previous inner product figures are apparent in these: the approximate doubling
for the depth L = 6 transformer can be visualized by the vertically changing opaque colors. Conversely, a
9Since these experiments are in the small alphabet size |Σ| = 4 regime, this task performs better than random guessing due to
inferential capabilities that are are powered by the high embedding dimension and do not require implementing a pointer-chasing
algorithm. We suspect that the “checkerboard” patterns are powered by this inference.
49
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
j
= 1, h = 1
= 1, h = 2
= 1, h = 3
= 1, h = 4
= 2, h = 1
= 2, h = 2
= 2, h = 3
= 2, h = 4
= 3, h = 1
= 3, h = 2
= 3, h = 3
= 3, h = 4
= 4, h = 1
= 4, h = 2
= 4, h = 3
= 4, h = 4
Layer , head h
0.32
0.17
0.1
0.01
0.01
0.32
0.32
0.15
0.08
0.05
0.03
0.02
0.01
0.01
0.04
0.19
0.08
0.08
0.04
0.03
0.01
0.01
0.93
0.77
0.05
0.11
0.02
0.02
0.52
0.04
0.14
0.04
0.04
0.02
0.02
0.01
0.01
0.01
0.8
0.07
0.94
0.01
0.19
0.18
0.15
0.08
0.05
0.02
0.01
0.23
0.09
0.15
0.05
0.04
0.02
0.01
0.18
0.77
0.03
0.8
0.17
0.01
0.93
0.01
0.87
0.04
A , h, findj n, 4 for depth-4 transformer and X
dom(hop4)
0.0
0.2
0.4
0.6
0.8
1.0
Figure 11: Plots of all inner products
Aℓ,h[T ∞
4,4], findj
10,4 for n = 10 samples X1, . . . , X10 ∈dom(hop4) for
the 4-layer transformer T ∞
4,4.
separation can be observed between the tasks where the depth L = 4 transformer performs well and has
“decisive” self-attention units deeper in the network and those where it does not.
Moreover, the figures (especially Figure 15) demonstrate that several self-attention units have a consistent
function among samples from the same task, while adapting in function to different hopk tasks. This is most
apparent in head h = 4 of layer ℓ= 6, where the self-attention head functions as find1, find3, find5 or find7
depending on the complexity of the task.
50
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
j
= 1, h = 1
= 1, h = 2
= 1, h = 3
= 1, h = 4
= 2, h = 1
= 2, h = 2
= 2, h = 3
= 2, h = 4
= 3, h = 1
= 3, h = 2
= 3, h = 3
= 3, h = 4
= 4, h = 1
= 4, h = 2
= 4, h = 3
= 4, h = 4
= 5, h = 1
= 5, h = 2
= 5, h = 3
= 5, h = 4
= 6, h = 1
= 6, h = 2
= 6, h = 3
= 6, h = 4
Layer , head h
0.33
0.33
0.16
0.09
0.05
0.03
0.02
0.01
0.01
0.34
0.01
0.32
0.24
0.08
0.05
0.02
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.88
0.06
0.01
0.08
0.22
0.06
0.06
0.02
0.01
0.01
0.92
0.01
0.93
0.93
0.26
0.06
0.07
0.03
0.03
0.03
0.02
0.02
0.02
0.03
0.03
0.03
0.03
0.03
0.03
0.02
0.21
0.14
0.13
0.06
0.04
0.02
0.01
0.01
0.01
0.01
0.93
0.9
0.14
0.11
0.08
0.05
0.03
0.02
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.07
0.79
0.06
0.93
0.04
0.81
0.04
0.03
0.89
0.01
0.02
0.9
0.01
0.02
0.07
0.77
0.03
0.02
0.01
0.12
0.01
0.65
0.04
0.03
0.01
0.14
0.74
0.13
0.01
0.01
0.01
0.01
0.04
0.82
0.02
A , h, findj n, 16 for depth-6 transformer and X
dom(hop16)
0.0
0.2
0.4
0.6
0.8
1.0
Figure 12: Plots of all inner products
Aℓ,h[T ∞
4,6], findj
10,16 for n = 10 samples X1, . . . , X10 ∈dom(hop16)
for the 6-layer transformer T ∞
4,6.
51
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
j
= 1, h = 1
= 1, h = 2
= 1, h = 3
= 1, h = 4
= 2, h = 1
= 2, h = 2
= 2, h = 3
= 2, h = 4
= 3, h = 1
= 3, h = 2
= 3, h = 3
= 3, h = 4
= 4, h = 1
= 4, h = 2
= 4, h = 3
= 4, h = 4
Layer , head h
0.32
0.23
0.13
0.01
0.01
0.33
0.21
0.09
0.05
0.03
0.02
0.01
0.01
0.01
0.04
0.22
0.1
0.11
0.06
0.05
0.03
0.02
0.01
0.01
0.92
0.02
0.64
0.05
0.19
0.03
0.05
0.01
0.01
0.01
0.51
0.05
0.17
0.05
0.07
0.03
0.03
0.02
0.02
0.01
0.01
0.01
0.01
0.46
0.29
0.08
0.01
0.68
0.21
0.06
0.01
0.05
0.06
0.15
0.08
0.15
0.07
0.09
0.04
0.03
0.02
0.01
0.01
0.11
0.12
0.1
0.11
0.08
0.06
0.04
0.03
0.02
0.01
0.01
0.02
0.14
0.25
0.21
0.1
0.03
0.01
0.01
0.01
0.01
0.07
0.01
0.16
0.02
0.17
0.03
0.11
0.02
0.05
0.01
0.02
0.61
0.32
0.11
0.03
0.01
0.01
0.03
0.21
0.31
0.01
0.17
0.01
0.06
0.01
A , h, findj n, 16 for depth-4 transformer and X
dom(hop16)
0.0
0.2
0.4
0.6
0.8
1.0
Figure 13: Plots of all inner products
Aℓ,h[T ∞
4,4], findj
10,16 for n = 10 samples X1, . . . , X10 ∈dom(hop16)
for the 4-layer transformer T ∞
4,4.
52
1 (0.00)
2 (0.00)
3 (0.01)
4 (0.01)
5 (0.03)
6 (0.05)
7 (0.09)
8 (0.17)
9 (0.21)
10 (0.26)
11 (0.27)
12 (0.30)
13 (0.33)
14 (0.32)
15 (0.32)
16 (0.32)
k (Error errk
n(Tn
5, 5))
= 1, h = 1
= 1, h = 2
= 1, h = 3
= 1, h = 4
= 2, h = 1
= 2, h = 2
= 2, h = 3
= 2, h = 4
= 3, h = 1
= 3, h = 2
= 3, h = 3
= 3, h = 4
= 4, h = 1
= 4, h = 2
= 4, h = 3
= 4, h = 4
Layer , head h
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
3
3
3
3
3
3
3
3
3
3
5
1
1
1
1
1
1
1
1
1
1
1
1
1
2
2
2
1
1
1
3
3
3
3
3
3
5
5
5
5
5
5
5
1
2
3
2
3
4
5
4
5
6
7
6
7
6
7
8
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
4
4
4
4
6
4
6
6
6
6
argmaxj A , h, findj n, k for depth-4 transformer and X
dom(hopk)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
j
Figure 14:
Plots of all the maximum inner products
Aℓ,h[T ∞
4,4], findj
n,k for n = 10 fixed samples
X1, . . . , X10 ∈dom(hopk) for each k ∈[16] for the 4-layer transformer T ∞
4,4.
The hue corresponds to
the index of the largest inner product j∗= arg maxj
Aℓ,h[T ∞
4,4], findj
n,k, while the opacity is determined by
the magnitude of the correlation.
53
1 (0.00)
2 (0.00)
3 (0.00)
4 (0.00)
5 (0.00)
6 (0.00)
7 (0.01)
8 (0.00)
9 (0.01)
10 (0.01)
11 (0.01)
12 (0.01)
13 (0.01)
14 (0.02)
15 (0.02)
16 (0.03)
k (Error errk
n(Tn
5, 7))
= 1, h = 1
= 1, h = 2
= 1, h = 3
= 1, h = 4
= 2, h = 1
= 2, h = 2
= 2, h = 3
= 2, h = 4
= 3, h = 1
= 3, h = 2
= 3, h = 3
= 3, h = 4
= 4, h = 1
= 4, h = 2
= 4, h = 3
= 4, h = 4
= 5, h = 1
= 5, h = 2
= 5, h = 3
= 5, h = 4
= 6, h = 1
= 6, h = 2
= 6, h = 3
= 6, h = 4
Layer , head h
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
3
2
1
3
3
3
3
3
3
3
3
3
3
3
3
3
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
3
3
3
3
3
3
5
5
5
5
5
1
1
1
1
1
1
1
1
1
3
3
3
3
3
3
3
1
1
1
1
1
1
1
3
3
3
3
3
3
3
3
3
1
2
1
2
3
2
3
4
5
4
5
6
7
6
7
8
2
1
2
3
2
3
4
5
4
5
6
7
6
7
8
9
2
1
1
1
1
1
1
1
1
3
3
3
3
5
5
5
2
1
1
1
1
3
3
3
3
5
5
5
5
7
7
7
argmaxj A , h, findj n, k for depth-6 transformer and X
dom(hopk)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
j
Figure 15:
Plots of all the maximum inner products
Aℓ,h[T ∞
4,6], findj
n,k for n = 10 fixed samples
X1, . . . , X10 ∈dom(hopk) for each k ∈[16] for the 6-layer transformer T ∞
4,6.
54
G.5
Finite-sample experiments (Empirical Claim G.3; Figures 16 to 19)
While most of our multi-hop experiments reside in the infinite-sample regime (where new samples are
generated for every batch), we also trained several transformers on ntrain ∈{1000, 3000} samples to evaluate
whether generalization is possible in this domain, especially when the number of model parameters far exceeds
the number of training samples. The two training set sizes expose a sharp threshold between two different
generalization modes: low accuracy due to overfitting for most models on most tasks when ntrain = 1000 and
high accuracy approaching the infinite-sample regime when ntrain = 3000.
Figure 16 compares the infinite-sample transformers T ∞
4,L with the 3000-sample models T 3000
4,L . 3000 training
samples are sufficient to obtain comparable (if slightly worse) generalization error rates across model depths
L and task complexities k. This supports a hypothesis that the existence of a small transformer that perfectly
fits the data enables larger transformers to actually realize such architectures in the over-parameterized
regime.
On the other hand, Figure 17 demonstrates that transformers trained on ntrain = 1000 samples suffer
poor performance on most tasks due to overfitting. While all models perform poorly on hopk sub-tasks for
large k, a depth-separation exists for simpler sub-tasks like hop3. This suggests a positive inductive bias of
deep transformers for simple compositional decision rules, which enables far better performance than other
models in the overfitting regime.
To investigate this gap in performance, we contrast the self-attention inner products of depth-4 T 1000
4,4
and
depth-6 T 1000
4,6
on the task hop3 in Figures 18 and 19. The 6-layer model obtains a far superior classification
error on the sub-task, and the interpretability plot establishes a plausible circuit it implements: It uses
self-attention heads with find1 functionality consecutively in layers 4, 5, and 6, which enables the robust
retrieval of find3 and hop3. On the other hand, the 4-layer plot exhibits poor performance and only has two
layers with find1 functionality; this justifies the relatively strong performance of T 1000
4,4
on hop2 and its poor
performance on hop3.
While neither model learns any kind of pointer-doubling construction, the 6-layer model is still able to
learn a simple construction of hop3 that the 4-layer model misses. The representational suitability of deeper
models to compositional reasoning may thus provide a favorable inductive bias for learning the task in a
setting with little data.
55
1
2
4
8
16
k
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Error: errk
n(T)
Evaluation of L-depth, 4-headed, 3000-sample intransformers on hopk
L = 2, n = 3000
L = 2, n =
L = 3, n = 3000
L = 3, n =
L = 4, n = 3000
L = 4, n =
L = 5, n = 3000
L = 5, n =
L = 6, n = 3000
L = 6, n =
Figure 16: Comparison between the errors errn
k(T n
4,L) of transformers trained in the infinite sample regime
(dashed line) and on ntrain = 3000 samples (solid line) on the multi-hop task, evaluated on n = 100 samples
per hopk task.
1
2
4
8
16
k
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Error: errk
n(T)
Evaluation of L-depth, 4-headed, 1000-sample intransformers on hopk
L = 2, n = 1000
L = 2, n =
L = 3, n = 1000
L = 3, n =
L = 4, n = 1000
L = 4, n =
L = 5, n = 1000
L = 5, n =
L = 6, n = 1000
L = 6, n =
Figure 17: Comparison between the errors errn
k(T n
4,L) of transformers trained in the infinite sample regime
(dashed line) and on ntrain = 1000 samples (solid line) on the multi-hop task, evaluated on n = 100 samples
per hopk task.
56
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
j
= 1, h = 1
= 1, h = 2
= 1, h = 3
= 1, h = 4
= 2, h = 1
= 2, h = 2
= 2, h = 3
= 2, h = 4
= 3, h = 1
= 3, h = 2
= 3, h = 3
= 3, h = 4
= 4, h = 1
= 4, h = 2
= 4, h = 3
= 4, h = 4
Layer , head h
0.15
0.09
0.1
0.08
0.07
0.06
0.05
0.05
0.04
0.04
0.03
0.03
0.03
0.02
0.02
0.02
0.31
0.06
0.04
0.03
0.02
0.02
0.02
0.02
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.04
0.04
0.03
0.03
0.03
0.02
0.02
0.02
0.02
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.32
0.07
0.05
0.04
0.03
0.03
0.02
0.02
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.22
0.1
0.05
0.04
0.03
0.02
0.01
0.01
0.01
0.09
0.12
0.07
0.09
0.06
0.06
0.05
0.05
0.05
0.04
0.04
0.04
0.03
0.03
0.03
0.03
0.28
0.11
0.12
0.08
0.07
0.05
0.05
0.04
0.03
0.03
0.02
0.02
0.02
0.02
0.02
0.02
0.31
0.13
0.11
0.06
0.05
0.03
0.02
0.02
0.02
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.19
0.13
0.12
0.09
0.07
0.05
0.04
0.03
0.02
0.02
0.01
0.01
0.01
0.01
0.05
0.2
0.06
0.07
0.03
0.02
0.01
0.01
0.01
0.01
0.88
0.06
0.01
0.86
0.07
0.01
0.85
0.09
0.01
0.61
0.04
0.12
0.02
0.03
0.01
0.01
0.33
0.06
0.17
0.06
0.08
0.04
0.04
0.03
0.02
0.01
0.01
0.01
0.03
0.4
0.02
0.09
0.02
0.03
0.01
0.01
A , h, findj n, 3 for depth-4 transformer and X
dom(hop3), trained on ntr = 1000 samples
0.0
0.2
0.4
0.6
0.8
1.0
Figure 18: Plots of all inner products
Aℓ,h[T 1000
4,4 ], findj
10,3 for n = 10 samples X1, . . . , X10 ∈dom(hop3)
for the 4-layer transformer T 1000
4,4 .
57
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
j
= 1, h = 1
= 1, h = 2
= 1, h = 3
= 1, h = 4
= 2, h = 1
= 2, h = 2
= 2, h = 3
= 2, h = 4
= 3, h = 1
= 3, h = 2
= 3, h = 3
= 3, h = 4
= 4, h = 1
= 4, h = 2
= 4, h = 3
= 4, h = 4
= 5, h = 1
= 5, h = 2
= 5, h = 3
= 5, h = 4
= 6, h = 1
= 6, h = 2
= 6, h = 3
= 6, h = 4
Layer , head h
0.33
0.04
0.03
0.02
0.02
0.01
0.01
0.01
0.01
0.01
0.01
0.24
0.14
0.11
0.09
0.07
0.06
0.05
0.04
0.04
0.03
0.03
0.02
0.02
0.02
0.02
0.02
0.26
0.09
0.07
0.06
0.05
0.05
0.04
0.03
0.02
0.02
0.02
0.02
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.18
0.12
0.09
0.06
0.05
0.04
0.03
0.03
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.32
0.13
0.12
0.06
0.05
0.04
0.03
0.02
0.02
0.02
0.01
0.01
0.01
0.01
0.01
0.01
0.26
0.15
0.09
0.06
0.04
0.02
0.02
0.01
0.01
0.01
0.16
0.1
0.11
0.08
0.07
0.06
0.05
0.05
0.04
0.04
0.03
0.03
0.03
0.03
0.02
0.02
0.15
0.14
0.04
0.03
0.01
0.01
0.25
0.09
0.06
0.02
0.01
0.01
0.01
0.01
0.01
0.01
0.32
0.13
0.11
0.06
0.05
0.03
0.02
0.02
0.02
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.34
0.13
0.11
0.08
0.06
0.05
0.04
0.03
0.02
0.02
0.02
0.02
0.01
0.01
0.01
0.01
0.01
0.93
0.01
0.92
0.02
0.13
0.19
0.09
0.08
0.04
0.03
0.02
0.01
0.01
0.01
0.42
0.08
0.1
0.03
0.03
0.02
0.01
0.01
0.01
0.93
0.02
0.91
0.04
0.9
0.04
0.92
0.03
0.11
0.07
0.16
0.07
0.09
0.05
0.04
0.03
0.02
0.02
0.01
0.01
0.01
0.01
0.91
0.03
0.12
0.55
0.04
0.1
0.02
0.02
0.01
0.01
0.01
A , h, findj n, 3 for depth-6 transformer and X
dom(hop3), trained on ntr = 1000 samples
0.0
0.2
0.4
0.6
0.8
1.0
Figure 19: Plots of all inner products
Aℓ,h[T 1000
4,6 ], findj
10,3 for n = 10 samples X1, . . . , X10 ∈dom(hop3)
for the 6-layer transformer T 1000
4,6 .
58
|