File size: 102,792 Bytes
8cf8144 |
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 |
\documentclass[12pt]{amsart}
\usepackage[margin=1in]{geometry}
\usepackage{graphicx} \usepackage[all]{xy}
\usepackage{schemabloc}
\usepackage{amssymb}
\usepackage{amsthm}
\usepackage{amsmath}
\usepackage{latexsym}
\usepackage[mathscr]{euscript}
\usepackage{pb-diagram}
\usepackage{latexsym}
\usepackage{epsfig}
\usepackage{graphics}
\usepackage{mathabx}
\newcommand\sep{\mathrm{sep}}
\newcommand\diam{\mathrm{diam}}
\newcommand{\beq}[1]{\begin{equation}\label{#1}}
\newcommand{\eeq}{\end{equation}}
\def\compcirc {\mbox{\hspace{.05cm}}\raisebox{.04cm}{\tiny {$\circ$ }}}
\newcommand\R{\mathbb{R}}
\newcommand\N{\mathbb{N}}
\newcommand\dmtwo[1]{\left(\begin{smallmatrix}
0 & {#1}\\
{#1} & 0
\end{smallmatrix}\right)}
\newcommand\dmthree[3]{\left(\begin{smallmatrix}
0 & {#1}&{#2}\\
{#1} & 0 & {#3}\\
{#2} & {#3}& 0
\end{smallmatrix}\right)}
\newcommand \cfunc{{\mathfrak C}}
\newcommand \hfunc{{\mathfrak H}}
\newcommand \rfunc{{\mathfrak R}}
\newcommand \func[1]{{\mathfrak #1}}
\newcommand \tfunc{{\mathfrak T}}
\newcommand \ifunc{{\mathfrak I}}
\newcommand\ob{\mathrm{ob}}
\newcommand\mor{\mathrm{Mor}}
\newcommand\im{\mathrm{Im}}
\newcommand \miso{\underline{\mathcal M}^{iso}}
\newcommand \mmon{\underline{\mathcal M}^{inj}}
\newcommand \mgen{\underline{\mathcal M}^{gen}}
\newcommand \many{\underline{\mathcal M}}
\newcommand \catc{\underline{\mathcal C}}
\newcommand \catp{\underline{\mathcal P}}
\newcommand \catq{\underline{\mathcal Q}}
\newcommand \cat[1]{\underline{\mathcal {#1}}}
\newcommand \cati{\underline{\mathcal I}}
\newcommand \catd{\underline{\mathcal D}}
\newcommand \catr{\underline{\mathbb E}}
\renewcommand{\theequation}{\thesection--\arabic{equation}}
\newtheorem{corollary}{Corollary}[section]
\newtheorem{theorem}{Theorem}[section]
\newtheorem{definition}{Definition}[section]
\newtheorem{example}{Example}[section]
\newtheorem{lemma}{Lemma}[section]
\newtheorem{obs}{Observation}[section]
\newtheorem{claim}{Claim}[section]
\newtheorem{remark}{Remark}[section]
\newtheorem{oproblem}{Open Problem}
\newtheorem{proposition}{Proposition}[section]
\newtheorem{assumption}{Assumption}[section]
\begin{document}
\title{Classifying clustering schemes}
\author{Gunnar Carlsson and Facundo M\'emoli}
\address{Department of Mathematics, Stanford University,
Stanford, CA 94305}
\thanks{This work is supported by DARPA grant HR0011-05-1-0007 and ONR grant N00014-09-1-0783.}
\email{$\{$gunnar,memoli$\}$@math.stanford.edu}
\date{November 2010}
\maketitle
\begin{abstract}
Many clustering schemes are defined by optimizing an objective
function defined on the partitions of the underlying set of a finite metric space.
In this paper, we construct a framework for studying what happens when we instead impose various
structural conditions on the clustering schemes, under the general heading of \emph{functoriality}.
Functoriality refers to the idea that one
should be able to compare the results of clustering algorithms as one
varies the data set, for example by adding points or by applying
functions to it. We show that within this framework, one can prove a
theorems analogous to one of J. Kleinberg \cite{kleinberg}, in which for example
one obtains an existence and uniqueness theorem instead of a
non-existence result.
We obtain a full
classification of all clustering schemes satisfying a condition we refer
to as excisiveness. The classification can be changed by varying the
notion of maps of finite metric spaces. The conditions occur naturally
when one considers clustering as the statistical version of the geometric
notion of connected components. By varying the degree of functoriality that one requires from the schemes it is possible to construct richer families of clustering schemes that exhibit sensitivity to density.
\end{abstract}
\section{Introduction}
Clustering techniques play a very central role in various parts of
data analysis. This type of methods can give important clues to the structure of data
sets, and therefore suggest results and hypotheses in the underlying
science. There are many interesting methods of clustering available,
which have been applied to good effect in dealing with many datasets
of interest, and are regarded as important methods in exploratory
data analysis.
Most methods begin with a data set equipped with a notion of distance or metric. Starting with this input, a method might select a partition of the data set as one which optimizes a choice of an objective function. Other methods such as single, average, and complete linkage define clusterings, or rather nested families of clusterings, using a \emph{linkage function} defined between clusters.
Desirable properties of clustering algorithms come from practitioners who have intuitive notions of what is a good clustering: ``they know it when they see it''. This is of course not satisfactory and a theoretical understanding needs to be developed. However, one thing this intuition reflects is the fact that density needs to be incorporated in the clustering procedures. Single linkage clustering, a procedure that enjoys several nice theoretical properties, is notorious for its insensitivity to density, which is manifested in the so called \emph{chaining effect} \cite{lance67general}.
Other methods such as average linkage, complete linkage \cite[Chapter 3]{clusteringref} and $k$-means share the property that they exhibit some sort of sensitivity to density, but are unstable in a sense which has been made theoretically precise \cite{clust-um} and are therefore not well supported by theory. We believe that this disconnect between theory and practice should not exist, and in particular, in \cite{multi-param} we have constructed a theoretically sound framework that incorporates density via the use of $2$-dimensional persistence ideas.
Given that there is relatively little theory surrounding clustering methods, the subject is regarded by many as consisting of a collection of ad hoc methods for which it is difficult to generate a coherent picture \cite{science_art}. One exception to this rule is a paper of J. Kleinberg \cite{kleinberg}, in which he proves a theorem which demonstrates the impossibility of constructing a clustering algorithm with three seemingly plausible properties. Kleinberg regards a clustering scheme as a map $\cfunc$ that assigns each finite metric space $(X,d_X)$ one of its partitions. His ``axioms'' are \begin{itemize}
\item \emph{Scale invariance}: $\cfunc(X,d_X)=\cfunc(X,\lambda\cdot d_X)$ for all $\lambda>0$;
\item \emph{Surjectivity}: for all partitions $P$ of $X$ there exists a metric $d_X$ on $X$ with $\cfunc(X,d_X)=P$; and
\item \emph{Consistency}: upon reducing distances between points in the same cluster of $X$ (as produced by $\cfunc$), and increasing distances between points in different clusters, the result of applying the method to the new metric does not change.
\end{itemize}
\begin{theorem}[\cite{kleinberg}]
There exists no clustering algorithm that simultaneously satisfies scale invariance, surjectivity and consistency.
\end{theorem}
Kleinberg also points out that his results shed light on the trade-offs one has
to make in choosing clustering algorithms. In this paper, we produce
a variation on this theme, which we believe also has implications for
how one thinks about and applies clustering algorithms.
In our earlier paper \cite{clust-um}, in the context of hierarchical methods, we have established a variant on Kleinberg's theorem, in which instead of impossibility, one actually proves an \emph{existence and uniqueness} result, again assuming three plausible properties, slightly different from Kleinberg's. In the present paper, we make one more step towards understanding the theoretical foundations of clustering methods by expanding on the theme of \cite{clust-um} and developing a context in which one can obtain much richer classification results, which include versions of a number of familiar methods, including analogues of the clique clustering methods familiar in network and graph problems. We now describe our context and results.
\subsection{Overview of our methods and results}
We regard clustering as the ``statistical version" of the geometric notion of constructing the connected components of a topological space. Recall (see, e.g. \cite{munkres}) that the path
components of a topological space $X$ are the equivalence classes of
points in the space under the equivalence relation $\sim _{path}$,
where, for $x,x' \in X$, we have $x \sim _{path} y$ if and only if
there is a continuous map $\varphi: [0,1] \rightarrow X$ so that
$\varphi (0) = x$ and $\varphi (1) = x'$. In other words, two points
in $X$ are in the same path component if they are connected by a
continuous path in $X$. We let $\pi _0 (X)$ denote the set of connected components of the space $X$. A crucial observation about $\pi _0$ is that it is a \textbf{functor}, i.e. that not only does one construct a set (of components) associated to every topological space, but additionally, to every continuous map $f : X \rightarrow Y$ one associates a map of sets $\pi _0(X) \rightarrow \pi _0 (Y)$. We take the point of view that this \textbf{functoriality} is a key property to export from topological spaces to the finite metric spaces which form the input to clustering algorithms. Functoriality has been a very useful framework
for the discussion of a variety of problems within mathematics over
the last few decades, see
\cite{maclane}.
\begin{remark}
The reader may not be familiar with the notion of functoriality, as presented in this paper. The fundamental idea behind functoriality is that it is crucial to study the maps or transformations between mathematical objects as well as the objects themselves. Interpreted at this level of generality, one can readily see the importance of this notion via a number of examples.
\begin{enumerate}
\item{{\bf Fourier analysis:} The key observation of Fourier analysis in its various forms is the the study of the symmetry groups of a space allows one to construct very useful and conceptually important bases for spaces of functions on the space. So, the standard Fourier basis for complex valued functions on the circle is constructed by selecting functions which satisfy certain transformation laws under the rotational symmetries. This is of course also true for, e.g., the spherical harmonic basis for the functions on $\Bbb{R}^3$. Functoriality can be regarded as a transformation law. }
\item{{\bf Normal forms for matrices:} The analysis of the conjugation action of groups of invertible matrices on themselves gives rise to normal forms of matrices, such as Jordan normal form. This means that the conjugation symmetries on spaces of matrices allows one to understand their organization}
\item{{\bf Algebraic topology:} In algebraic topology, one associates to topological spaces algebraic objects called {\em homology groups}. It is critical for applications of these groups, and for computing them, that they obey a transformation law for any continuous map of spaces. This transformation law turns out to be very powerful, in that it is a key factor in making a small axiomatic description of the homology groups. Just as transformation laws for symmetries determine the Fourier basis, so the functoriality determines the homology groups. }
\item{{\bf Galois theory:} In Galois theory, groups of symmetries allow one to describe the sets of solutions of algebraic equations over fields, such as the rational numbers. The observation that equations were associated to symmetry groups revolutionized number theory, and to this day the study of various kinds of transformation laws in modular forms continues to have a powerful impact on the field. }
\end{enumerate}
\end{remark}
The point of the present paper is to illustrate that functoriality allows a very useful framework for classifying large families of clustering schemes. This framework will include clustering schemes which are sensitive to effects by density within the domain data set. Moreover, the functoriality of the construction of path connected components is a key feature in all aspects of topology, and the functoriality of clustering constructions should play a similar role in the computational topology of finite data sets.
Then, we require that a standard\footnote{As opposed to hierarchical.} clustering algorithm $\cfunc$ be a rule which assigns to every finite metric space $X$ a set of clusters $\cfunc (X)$, subject to additional the requirement that every map $f: X \rightarrow Y$ induce a map of sets $\cfunc(f) : \cfunc (X) \rightarrow \cfunc (Y)$.
The question is what the replacement for continuous maps (i.e. \emph{morphisms} of topological spaces) should be, and we find that there are many choices. We consider three nested classes of maps, namely isometries (called $iso$), distance non-increasing maps (called $gen$, for general), and distance non-increasing maps which are injections on the sets of points (called $inj$), with the hierarchy being given by $iso\subset inj\subset gen.$ We find that functoriality with respect to $gen$ is quite restrictive, in that it singles out single linkage clustering as the only functor which satisfies the $gen$ functoriality requirement. On the other hand, functoriality with respect to $iso$ appears not to be restrictive enough, in that it permits the specification of an arbitrary clustering on every isometry class of finite metric spaces.
One of the features of our classification is the proof that for the $gen$ and $inj$ cases, a certain implicitly defined property of clustering schemes called \textbf{excisiveness}, which amounts to requiring that the clustering method is idempotent on each of the blocks of the partition it generates, is \emph{equivalent} to the existence of a certain kind of \emph{explicit generative model} for the scheme.
Mathematical results which prove the existence of a generative model given that some externally defined properties are satisfied are very powerful in mathematics. For example, the result that a system satisfying the axioms for a vector space over a field $\mathbb{F}$ admits a basis is very useful, as is the fact that a graph with no odd order cycles is bipartite.
The equivalence between excisiveness and the existence of a generative model for the clustering scheme has the interpretation that excisive schemes are parametrized by sets of ``test metric spaces", which are those connected into a single cluster by the scheme. The method associated to such a family is then defined using a criterion which asserts that two points $x$ and $x'$ in a metric space $X$ belong to the same cluster if there is a sequence of points $\{ x_0, x_1, \ldots , x_n \}$ with $x_0 = x$, $x_n = x'$, and so that for each pair $x_i, x_{i+1}$ there is a morphism in $gen$ or $inj$ from one of the test metric spaces into $X$ with $x_i$ and $x_{i+1}$ in its image.
A key feature of the choice of morphisms in $inj$ is that (since they are injective) this permits one to take account of \emph{density} in a certain sense. If one considers functoriality with respect to $gen$, one finds that the morphisms might collapse a large dense collection of points into a single point, which makes it impossible to take density into account in building the clusters. Our classification of functors based on $inj$ includes analogues of ``clique clustering" \cite{hungarian-clique} and the DBSCAN algorithm \cite{dbscan}, and such methods clearly have the effect of taking a version of density into account.
Whereas our classification of $gen$ functorial schemes yields that single linkage is the unique possible scheme, which turns out to be excisive, we find large classes of non-excisive clustering schemes in $inj$. Our construction of such families relies on a study of the metric invariants that are well behaved under $inj$ morphisms. Finally, it turns out that all schemes in $inj$ that admit a \emph{finite} generative model arise as the composition of single linkage clustering with an application that changes the underlying metric of the input metric space. This result has clear practical consequences which we discuss.
We believe that functoriality, incorporating as it does the idea that clustering should not operate only on finite metric spaces as isolated objects, but rather on the metric space together with other metric spaces which are related to it via morphisms, is a very desirable property. We can point to its success within topology as an extremely powerful tool which pins down many constructions as the unique solution to existence problems, and which is the key to most applications
of topological methods. It is also true that functoriality is necessary in order to extend the methods of clustering to higher order connectivity information with computational topology of finite data sets \cite{carlsson-09}.
In this paper we demonstrate that the use of restricted notions of functoriality allows us to study clustering schemes which take proxies for density into account. This is important because it is clear that some of the value of alternate schemes such as average and complete linkage (which are not functorial) is due precisely to their ability to do this. Therefore, some of the methods we study will not suffer from the problems inherent in single linkage clustering, such as chaining \cite[Section 7.4]{jardine-sibson} \cite{multi-param}.
\subsection{Organization of the paper} In Section
\ref{sec:prelim} we discuss basic background concepts that we use in our presentation. Section \ref{cats} introduces the concept of a category and gives many examples and constructions we use in the paper. The concept of a functor and functoriality are discussed in Section \ref{sec:functors}.
We present our main characterization results for standard clustering methods in Section
\ref{results-sc}. We study hierarchical methods in Section \ref{results-hc}. Concluding remarks are given in Section \ref{various}.
\section{Preliminaries}
\label{sec:prelim}
For a finite set $X$ we denote by $\mathcal{P}(X)$ the set of all partitions of $X$. For each $k\in\N$, $\mathcal{M}_k$ denotes the collection of all finite metric spaces with cardinality $k$ and $\mathcal{M}=\bigcup_{k\in\N} \mathcal{M}_k$ denotes the collection of all finite metric spaces. For $(X,d_X)\in\mathcal{M}$ let $\sep(X,d_X):=\min_{x\neq x'}d_X(x,x')$ be the \emph{separation} of $X$, $\diam(X,d_X):=\max_{x,x'\in X}d_X(x,x')$ be the \emph{diameter} of $X$, and for $\lambda>0$ we denote by $\lambda\cdot X$ the metric space $(X,\lambda\cdot d_X)$. For $\delta>0$ and $k\geq 2$ we let $\Delta_k(\delta)$ denote the metric space with $k$ points whose inter-point distances equal $\delta$.
\begin{definition}\label{def-equiv-rips} On $(X,d_X)\in\mathcal{M}$, for each $\delta\geq 0$, we define the equivalence relation $\sim _\delta$, where $x \sim _\delta x^{\prime}$ if and only if there is a sequence $x_0, x_1, \ldots , x_k \in X$ so that $x_0 = x, x_k = x^{\prime}$, and $d_X(x_i, x_{i+1}) \leq \delta$ for all $i$. We denote by $[x]_\delta$ the equivalence class of $x\in X$ under $\sim_\delta$.
\end{definition}
The following lemma follows directly from the definition of $\sim_\delta$.
\begin{lemma}\label{prop:props-uaster} Let $(X,d_X)\in\mathcal{M}$ and $\delta\geq 0$. Then,
$[x]_\delta\neq [x']_\delta$ for some $x,x'\in X$ if and only if { $$d_X(a,a')>\delta,\,\,\mbox{for all $a\in [x]_\delta$ and $a'\in [x']_\delta$}.$$}
\end{lemma}
Let $X$ be any finite set and $W_X:X\times X\rightarrow \R_+$ be a symmetric function s.t. $W_X(x,x)=0$ for all $x\in X$. Then, the \emph{maximal sub-dominant ultrametric} relative to $W_X$ is given by
{\beq{eq:U}\mathcal{U}(W_X)(x,x')=\min\Big\{\max_{i}W_X(x_i,x_{i+1}),\,\{x_i\}_{i=0}^k\,\mbox{s.t. $x_0=x$ and $x_k=x'$}\Big\}
.\eeq}
This construction and its correctness are standard \cite[\S 6.4.3]{handbook-um}.
We define objects which will encode
the notion of ``multiscale'' or ``multiresolution'' partitions.
\begin{definition}\label{def:persistent-sets}
A {\em persistent set} is a pair $(X, \theta_X)$, where $X$ is a finite
set, and $\theta_X$ is a function from the non-negative real line $[0,+
\infty)$ to ${\mathcal P}(X)$ so that the following
properties hold.
\begin{enumerate}
\item{ If $r \leq s$, then $\theta_X (r)$ refines $\theta_X(s)$. }
\item{For any $r$, there is a number $\epsilon > 0$ so that $\theta_X (r^{\prime}) = \theta_X (r)$ for all $r^\prime\in[r,r+\epsilon]$.} \end{enumerate}
If in addition there exists $t>0$ s.t. $\theta_X(t)$ consists of the
single block partition for all $r\geq t$, then we say that
$(X,\theta_X)$ is a \emph{dendrogram}.\footnote{In the paper we will be
using the word dendrogram to refer both to the object defined here and
to the standard graphical representation of them.}
\end{definition}
\begin{example}\label{rips} Let $(X,d_X)$ be a finite metric space. Then we can associate to $(X,d_X)$ the persistent set whose underlying set is $X$, and where for each $r\geq 0$, blocks of the partition $\theta_X (r)$ consist of the equivalence classes under the equivalence relation $\sim _r$.
\end{example}
\begin{example} \label{hierarchical} Here we consider the family
of Agglomerative Hierarchical clustering techniques,
\cite{clusteringref}. We (re)define these by the recursive
procedure described next. Let $X=\{x_1,\ldots,x_n\}$ and let
$\mathcal L$ denote a family of \emph{linkage functions},
i.e. functions which one uses for defining the distance between
two clusters. Fix $l\in {\mathcal L}$. For each $R>0$ consider the
equivalence relation $\sim_{l,R}$ on blocks of a partition
$\Pi\in{\mathcal P}(X)$, given by \mbox{${\mathcal B}\sim_{l,R}
{\mathcal B}'$} if and only if there is a sequence of blocks
${\mathcal B}={\mathcal B}_1,\ldots,{\mathcal B}_s={\mathcal B}'$
in $\Pi$ with $l({\mathcal B}_k,{\mathcal B}_{k+1})\leq R$ for
$k=1,\ldots,s-1$. Consider the sequences $r_1,r_2,\ldots\in
[0,\infty)$ and $\Theta_1,\Theta_2,\ldots\in \mathcal{P}(X)$ given
by $\Theta_1:=\{x_1,\ldots,x_n\}$ and for $i\geq 1$,
$\Theta_{i+1}=\Theta_i/\sim_{l,r_i}$ where $r_i:=
\min\{l({\mathcal B},{\mathcal B}'),\,{\mathcal B},{\mathcal
B}'\in\Theta_i,\,{\mathcal B}\neq {\mathcal B}'\}.$ Finally, we
define $\theta^l:[0,\infty)\rightarrow \mathcal{P}(X)$ by
$r\mapsto \theta^l(r):=\Theta_{i(r)}$ where $i(r):=\max\{i|r_i\leq
r\}$.
Standard choices for $l$ are \emph{single linkage}:
$$l({\mathcal B},{\mathcal B}')=\min_{x\in {\mathcal B}}\min_{x'\in
{\mathcal B}'}d_X(x,x');$$ \emph{complete linkage} $$l({\mathcal
B},{\mathcal B}')=\max_{x\in {\mathcal B}}\max_{x'\in {\mathcal
B}'}d_X(x,x');$$ and \emph{average linkage}: $$l({\mathcal
B},{\mathcal B}')=\frac{\sum_{x\in {\mathcal B}}\sum_{x'\in
{\mathcal B}'}d_X(x,x')}{|{\mathcal B}| \cdot |{\mathcal B}'|}.$$
It is easily verified that the notion discussed in Example
\ref{rips} is \emph{equivalent} to $\theta^l$ when $l$ is the
single linkage function. Note that, unlike the usual definition of
agglomerative hierarchical clustering, at each step of the
inductive definition we allow for more than two clusters to be
merged.
\end{example}
\section{Categories}\label{cats}
\subsection{Definitions and Examples}
In this section, we will give a brief description of the theory of
categories and functors, which will be the framework in which we state
the constraints that we require of our clustering algorithms. An
excellent reference for these ideas is \cite{maclane}.
Categories are useful mathematical constructs that encode the nature
of certain objects of interest together with a set of
admissible maps between them. This formalism is
extremely useful for studying classes of mathematical objects which
share a common structure, such as sets, groups, vector spaces, or
topological spaces. The definition is as follows.
\begin{definition} A \textbf{category} $\underline{C}$ consists of:
\begin{itemize}
\item{ A collection of \textbf{objects} $\ob(\underline{C})$ (e.g.
sets, groups, vector spaces, etc.)}
\item{ For each pair of objects $X,Y \in \ob(\underline{C})$, a
set\\
$\mor_{\underline{C}} (X,Y)$, the \textbf{morphisms} from $X$ to $Y$
(e.g. maps of sets from $X$ to $Y$, homomorphisms of groups
from $X$ to $Y$, linear transformations from $X$ to $Y$,
etc. respectively)}
\item{ Composition operations:\\ $\compcirc :
\mor_{\underline{C}} (X,Y) \times \mor_{\underline{C}} (Y,Z)
\rightarrow \mor_{\underline{C}} (X,Z)$, corresponding to
\textbf{composition} of set maps, group homomorphisms, linear
transformations, etc. }
\item{For each object $X \in \underline{C}$, a distinguished element $\textrm{id}_X\in \mor_{\underline{C}}(X,X)$, called the \textbf{identity} morphism.}
\end{itemize}
The composition is assumed to be associative in the obvious sense, and for any $f \in \mor_{\underline{C}} (X,Y)$, it is assumed that $\textrm{id}_Y \compcirc f = f$ and $f \compcirc \textrm{id}_X = f$.
\end{definition}
\begin{example}[The category of sets]
Denote by $\underline{Sets}$ the category whose objects are all sets, and the morphisms between two sets $A$ and $B$ are all the functions from $A$ to $B$.
\end{example}
A morphism $f:X \rightarrow Y$ which has a two sided inverse $g:
Y\rightarrow X$, so that $f \compcirc g = \textrm{id}_Y$ and $g \compcirc f =
\textrm{id}_X$, is called an {\bf isomorphism}. Two objects which are
isomorphic are intuitively thought of as ``structurally
indistinguishable" in the sense that they are identical except for
naming or choice of coordinates. For example, in the category of
sets, the sets $\{ 1,2,3 \}$ and $\{ A,B,C \}$ are isomorphic, since
they are identical except for the choice made in labelling the elements.
\begin{example}
Here we consider four very simple categories, that we are going to refer to as $\underline{0},\underline{1},\underline{2}$ and $\underline{3}$ respectively.
\begin{itemize}
\item The category $\underline{0}$ has $ob(\underline{0})=\emptyset$ and all the conditions in the definition above are trivially satisfied.
\item Consider the category $\underline{1}$ with exactly one object $A$ and one morphism: $\mor_{\underline{1}}(A,A)=f$. It follows that $f$ must be the identity morphism $id_A$. This is represented graphically as follows:
$$\xymatrix{
A \ar@(ur,dr)[]|{id_A}}$$
\item The category $\underline{2}$ has exactly two objects $A$ and $B$ and three morphisms: the identities from $A$ to $A$ and from $B$ to $B$ and exactly one morphism in $\mor_{\underline{2}}(A,B)$:
$$\xymatrix{
A \ar@(ul,dl)[]|{id_A} \ar[rr]|f
&& B \ar@(ur,dr)[]|{id_B} }$$
\item Finally, the category $\underline{3}$ has exactly three objects $A$, $B$ and $C$ and six morphisms: the identities from $A$ to $A$, from $B$ to $B$ and $C$ to $C$, and three more morphisms, $\mor_{\underline{3}}(A,B)=f$, $\mor_{\underline{3}}(B,C)=g$ and $\mor_{\underline{3}}(A,C)=h$:
$$\xymatrix{
A \ar@(ul,dl)[]|{id_A} \ar[rr]|f \ar[rd]|h && B \ar@(ur,dr)[]|{id_B} \ar[ld]|g\\
& C \ar@(dl,dr)[]|{id_C}
}
$$
\noindent
Now, note that in order to satisfy the composition rule one must have $h=g\circ f$.
\end{itemize}
\end{example}
\begin{example}
A more surprising example of a category is given by the following construction. A \emph{monoid} $(M,\ast,e)$ consists of a set $M$ equipped with a binary operation $\ast :M\times M\rightarrow M$ s.t. $(A\ast B)\ast C = A\ast (B\ast C)$ for all $A,B,C\in M$ and an element $e\in M$ s.t. $e\ast A = A\ast e =A$ for all $A\in M$. A monoid can be made into a category $\underline{M}$ with a single object $O$ and $\mor_{\underline{M}}(O,O)=M$ and $id_O$ being represented by $e$. The composition of morphisms is induced by the operation $\ast$.
\end{example}
\begin{example}
Let $\mathbb{E}$ denote the extended positive real line $[0,+\infty]$. We construct the category $\catr$ with $\ob(\catr)=\mathbb{E}$ and $\mor_{\catr}(a,b) = \{(a,b)\}$ if $a\geq b$, and $\mor_{\catr}(a,b)=\emptyset$. The composition is given by $(a,b)\compcirc (b,c) = (a,c)$.
\end{example}
Now we discuss the core constructions for this paper.
\begin{definition}[\bf $\catc$, a category of outputs of standard clustering schemes]{Let $Y$ be a finite set, $P_Y\in \mathcal{P}(Y)$, and $f:X\rightarrow Y$ be a set map. We define $f^*(P_Y)$ to be the partition of $X$ whose blocks are the sets $f^{-1}({B})$ where ${B}$ ranges over the blocks of $P_Y$.
We construct the category $\catc$ of outputs of standard
clustering algorithms with $\ob(\catc)$ equal to all
possible pairs $(X,P_X)$ where $X$ is a finite set and $P_X$ is a partition of $X$: $P_X\in {\mathcal P}(X)$. For objects $(X,P_X)$ and $(Y,P_Y)$ one sets $\mor_{\catc}\big((X,P_X),(Y,P_Y)\big)$ to be the set of all maps $f:X\rightarrow Y$ with the property that $P_X$ is a refinement of $f^*(P_Y)$.
}
\end{definition}
\begin{example}
Let $X$ be any finite set, $Y=\{a,b\}$ a set with two elements, and $P_X$ a partition of $X$. Assume first that $P_Y=\{\{a\},\{b\}\}$ and let $f:X\rightarrow Y$ be any map. Then, in order for $f$ to be a morphism in $\mor_{\catc}\big((X,P_X),(Y,P_Y)\big)$ it is necessary that $x$ and $x'$ be in different blocks of $P_X$ whenever $f(x)\neq f(x')$. Assume now that $P_Y = \{a,b\}$ and $g:Y\rightarrow X$. Then, the condition that $g\in\mor_{\catc}\big((Y,P_Y),(X,P_X)\big)$ requires that $g(a)$ and $g(b)$ be in the same block of $P_X$.
\end{example}
We will also construct a category of \emph{persistent sets}, which will constitute the output of hierarchical clustering functors.
\begin{definition}[\bf $\catp$, a category of outputs of hierarchical clustering schemes]\label{definition:P}
{ Let $(X, \theta_X), (Y, \theta_Y)$ be persistent sets. A
map of sets $f: X \rightarrow Y$ is said to be {\em persistence
preserving} if for each $r \in \mathbb{R}$, we have that $\theta_X (r)$
is a refinement of $ f^*(\theta_Y(r))$. We define a category $\underline{{\mathcal P}}$
whose objects are persistent sets, and where $\mor_{\underline{{\mathcal
P}}}((X,\theta_X), (Y, \theta_Y))$ consists of the set maps from $X$ to $Y$
which are persistence preserving. }
\end{definition}
It is easily verified that the
composite of persistence preserving maps is persistence preserving,
and that any identity map is persistence preserving. A simple example is shown in Figure
\ref{fig:pers-preserv}.
\begin{figure}[htb]
\centering
\includegraphics[width=1\linewidth]{persist-preserv.eps}
\caption{Two persistent sets $(X,\theta_X)$ and $(Y,\theta_Y)$
represented by their dendrograms. On the left one defined in
the set $X = \{A,B,C\}$ and on the right one defined on the
set $Y=\{A',B',C'\}$. Consider the given set map
$f:X\rightarrow Y$. Then we see that $f$ is persistence
preserving since for each $r\geq 0$, the partition $\theta_X(r)$
is a refinement of $f^*(\theta_Y(r))$. Indeed, there are three
interesting ranges of values of $r$. Pick for example $r$ like
in the orange shaded area: $r\in[1,2)$. Then $\theta_Y(r) =
\{\{A',B'\},\{C'\}\}$ and hence $f^*(\theta_Y(r)) =
\{f^{-1}(\{A',B'\}),\{f^{-1}(C')\}\} = \{\{A,B\},\{C\}\}$
which is indeed refined by
$\theta_X(r)=\{\{A\},\{B\},\{C\}\}$. One proceeds similarly for
the other two cases.} \label{fig:pers-preserv}
\end{figure}
\subsection{Three categories of finite metric spaces.}
{ We will describe three categories $\miso$, $\mmon$, and $\mgen$, whose
collections of objects will all consist of the collection of finite
metric spaces $\mathcal{M}$. For $(X,d_X)$ and $(Y, d_Y)$ in $\mathcal{M}$, a map $f: X \rightarrow Y$ is said to be {\bf distance
non increasing} if for all $x,x^{\prime} \in X$, we have
$d_Y(f(x),f(x^{\prime})) \leq d_X(x,x^{\prime})$. It is easy to check
that composition of distance non-increasing maps are also distance
non-increasing, and it is also clear that $\textrm{id}_X$ is always distance
non-increasing. We therefore have the category $\mgen$, whose objects are finite metric spaces, and so that for
any objects $X$ and $Y$, $\mor_{\mgen}(X,Y)$ is
the set of distance non-increasing maps from $X$ to $Y$, cf.
\cite{isbell} for another use of this class of maps. It is clear that compositions
of injective maps are injective, and that all identity maps are injective, so we
have the new category $\mmon$, in which
$\mor_{\mmon}(X,Y)$ consists of the \textbf{injective distance
non-increasing maps}. Finally, if $(X,d_X)$ and $(Y,d_Y)$ are finite
metric spaces, $f:X
\rightarrow Y$ is an {\bf isometry} if $f$ is bijective and $d_Y(f(x),
f(x^{\prime})) = d_X(x,x^{\prime})$ for all $x$ and $x^{\prime}$. It
is clear that as above, one can form a category $\miso$ whose objects are finite metric spaces and whose morphisms
are the isometries. Furthermore, one has inclusions
\begin{equation}\label{eq:inclu-cats}
\miso\subseteq \mmon
\subseteq \mgen
\end{equation}
of subcategories (defined as in \cite{maclane}). Note that although
the inclusions are bijections on object sets, they are proper
inclusions on morphism sets, i.e. in general they are not surjective.
}
\begin{remark}\label{remark:mgen}
The category $\mgen$ is special in that for any pair of finite metric spaces $X$ and $Y$, \mbox{$\mor_{\mgen}(X,Y)\neq \emptyset$}. Indeed, pick $y_0\in Y$ and define $\phi:X\rightarrow Y$ by $x\mapsto y_0$ for all $x\in X$. Clearly, $\phi\in\mor_{\mgen}(X,Y)$. This is not the case for $\mmon$ since in order for $\mor_{\mmon}(X,Y)\neq \emptyset$ to hold it is necessary (but not sufficient in general) that $|Y|\geq |X|$.
\end{remark}
\section{Functors and functoriality.} \label{sec:functors}
Next we introduce the key concept in our discussion, that of a
\emph{functor}. We give the formal definition first, and several examples will appear as different constructions that we use in the paper.
\begin{definition}[\bf Functor]\label{def:functor}
Let $\underline{C}$ and $\underline{D}$ be categories. Then a {\em functor} from $\underline{C}$ to $\underline{D}$ consists of:
\begin{itemize}
\item{A map of sets $F:\ob(\underline{C}) \rightarrow
\ob(\underline{D})$.}
\item{For every pair of objects $X,Y \in \underline{C}$ a map
of sets $\Phi(X,Y):\mor_{\underline{C}}(X,Y) \rightarrow
\mor_{\underline{D}}(FX,FY)$ so that
\begin{enumerate}
\item{ $\Phi(X,X)(\textrm{id}_X) = \textrm{id}_{F(X)}$ for all $X \in
\ob(\underline{C})$, and}
\item{$ \Phi (X,Z)(g \compcirc f) = \Phi (Y,Z)(g)\compcirc \Phi (X,Y)(f)$ for all $f \in \mor_{\underline{C}}(X,Y)$ and $g \in \mor_{\underline{C}}(Y,Z)$.}
\end{enumerate} }
\end{itemize}
Given a category $\underline{C}$, an \textbf{endofunctor} on $\underline{C}$ is any functor $F:\underline{C}\rightarrow \underline{C}.$
\end{definition}
\begin{remark}
In the interest of clarity, we will always refer to the pair $(F,\Phi)$ with a
single letter $F$. See diagram (\ref{eq:diagram-general}) below for an example.
\end{remark}
\begin{example}{\bf (Forgetful functors)} When one has two categories $\underline{C}$ and $\underline{D}$, where the objects in $\underline{C}$ are objects in $\underline{D}$ equipped with some additional structure and the morphisms in $\underline{C}$ are simply the morphisms in $\underline{D}$ which preserve that structure, then we obtain the ``forgetful functor" from $\underline{C}$ to $\underline{D}$, which carries the object in $\underline{C}$ to the same object in $\underline{C}$, but regarded without the additional structure. For example, a group can be regarded as a set with the additional structure of multiplication and inverse maps, and the group homomorphisms are simply the set maps which respect that structure. Accordingly, we have the functor from the category of groups to the category of sets which ``forgets the multiplication and inverse". Similarly, we have the forgetful functor from $\catc$ to the category of sets, which forgets the presence of $P_X$ in the output set $(X, P_X)$.
\end{example}
\begin{example}
The inclusions $ \miso\subseteq \mmon \subseteq \mgen $ are both functors.
\end{example}
\begin{example}[{Scaling functor}]\label{scaling} For any $\lambda >0$ we define an endofunctor $\sigma _{\lambda} : \mgen
\rightarrow \mgen$ on objects by
$ \sigma _{\lambda}(X,d_X) = (X,\lambda \cdot d_X)
$
and on morphisms by $\sigma _{\lambda} (f) = f$. One easily verifies that if $f$ satisfies the conditions for being a morphism in $\mgen$ from $(X,d_X)$ to $(Y,d_Y)$, then it readily satisfies the conditions of being a morphism from $(X,\lambda \cdot d_X)$ to $(Y, \lambda \cdot d_Y)$. Clearly, $\sigma_\lambda$ can also be regarded as an endofunctor in $\miso$ and $\mmon.$
Similarly, we define a functor $s_{\lambda} : \underline{{\mathcal P}} \rightarrow \underline{{\mathcal P}}$ by setting $s_{\lambda}(X, \theta _X) = (X, \theta_X ^{\lambda})$, where
$\theta_X ^{\lambda}(r) = \theta_X (\frac{r}{\lambda})$.
\end{example}
\subsection{Clustering algorithms as functors}
The notion of categories, functors and functoriality provide
useful framework for studying algorithms. One first defines a class of
input objects $\underline{\mathcal{I}}$ and a class of output objects $\underline{\mathcal{O}}$. Moreover, one
associates to each of these classes a class of natural maps, the morphisms, between
objects. For the problem of HC for example, the input class is the
set of finite metric spaces and the output class is that of
dendrograms. An algorithm is to be regarded as a functor between a category of input objects and a category of output objects.
An algorithm will therefore be a procedure that assigns to each $I\in\underline{\mathcal
I}$ an output $O_I\in\underline{{\mathcal O}}$ with the further property that
it respects relations between objects in the following sense. Assume
$I,I'\in \underline{\mathcal I}$ such that there is natural map
$f:I\rightarrow I'$. Then, the algorithm has to have the property
that the relation between $O_I$ and $O_{I'}$ has to be represented
by a natural map for output objects.
\begin{example}
Assume that $\cat{I}$ is such that $\mor_{\cat{I}}(X,Y)=\emptyset$ for all $X,Y\in \cat{I}$ with $X\neq Y$. In this case, since there are no morphisms between input objects any functor $\func{A}:\cat{I}\rightarrow \cat{O}$ can be specified arbitrarily on each $X\in\cat{O}$.
\end{example}
We view any given clustering scheme as a procedure which takes as input a
finite metric space $(X,d_X)$, and delivers as output either an object in $\catc$ or $\catp$:
\begin{itemize}
\item \textbf{Standard clustering}: a pair $(X,P_X)$ where $P_X$ is a partition of $X$. Such a pair is an object in the category $\catc$.
\item \textbf{Hierarchical clustering}: a pair $(X,\theta_X)$ where $\theta_X$ is a persistent set over $X$. Such a pair is an object in the category $\catp$.
\end{itemize}
The concept of \textbf{functoriality} refers to the additional condition
that the clustering procedure should map a pair of input objects into a pair
of output objects in a manner which is consistent with respect
to the morphisms attached to the input and output spaces. When this
happens, we say that the clustering scheme is \textbf{functorial}.
This notion of consistency is made precise in Definition
\ref{def:functor} and described by diagram (\ref{eq:diagram-general}). Let $\many$ stand for any of $\mgen$, $\mmon$ or $\miso$.
According to Definition \ref{def:functor}, in order to view a standard
clustering scheme as a functor $\cfunc:\many\rightarrow \catc$ we need to specify:
\begin{itemize}
\item[(1)] how it maps
objects of $\many$ (finite metric spaces)
into objects of $\catc$, and
\item[(2)]
how a morphism $f:(X,d_X)\rightarrow (Y,d_Y)$ between two
objects $(X,d_X)$ and $(Y,d_Y)$ in the input category
$\many$ induces a map in the output category
$\catc$, see diagram (\ref{eq:diagram-general}).
\end{itemize}
\begin{equation}\label{eq:diagram-general}
\xymatrix{
(X,d_X) \ar[d]^{\cfunc} \ar[r]^{f} & (Y,d_Y) \ar[d]^{\cfunc} \\
(X,P_X) \ar[r]^{\cfunc(f)} & (Y,P_Y)}
\end{equation}
Similarly, in order to view a hierarchical clustering scheme as a functor $\hfunc:\many\rightarrow \catp$ we need to specify:
\begin{itemize}
\item[(1)] how it maps
objects of $\many$ (finite metric spaces)
into objects of $\catp$, and
\item[(2)]
how a morphism $f:(X,d_X)\rightarrow (Y,d_Y)$ between two
objects $(X,d_X)$ and $(Y,d_Y)$ in the input category
$\many$ induces a map in the output category
$\catp$, see diagram (\ref{eq:diagram-general}).
\end{itemize}
\begin{equation}\label{eq:diagram-general}
\xymatrix{
(X,d_X) \ar[d]^{\hfunc} \ar[r]^{f} & (Y,d_Y) \ar[d]^{\hfunc} \\
(X,\theta_X) \ar[r]^{\hfunc(f)} & (Y,\theta_Y)}
\end{equation}
Precise constructions will be discussed in Sections \ref{results-sc} and \ref{results-hc}, where we study different clustering algorithms, both standard or flat, and hierarchical, using the idea of
functoriality.
We have 3 possible ``input'' categories ordered by
inclusion (\ref{eq:inclu-cats}). The idea is that studying
functoriality over a larger category will be more stringent/demanding
than requiring functoriality over a smaller one. We will consider
different clustering algorithms and study whether they are functorial
over our choice of the input category. The least demanding one, $\underline{\mathcal
M}^{iso}$ basically enforces that clustering schemes are not dependent on the way points are labeled. We will prove uniqueness results for functoriality over the most stringent category
$\underline{\mathcal M}^{gen}$, and finally we study how relaxing the
conditions imposed by the morphisms in $\underline{\mathcal M}^{gen}$,
namely, by restricting ourselves to the smaller but intermediate
category $\underline{\mathcal M}^{mon}$, we permit more functorial
clustering algorithms.
\section{Invariants of metric spaces}\label{sec:invariants}
Now we consider maps from a metric space into the ordered extended positive real line $\catr$ that behave well under morphisms of the categories $\miso$, $\mmon$, and $\mgen$, respectively. Such maps will be used later on for defining certain clustering functors, and the richness of such maps, or lack thereof, signals the degree to which these clustering functors exhibit different interesting behaviors.
\begin{definition}\label{def:invariant}
Let $\many$ be any of $\mgen$, $\mmon$ or $\miso$, then an \textbf{invariant} is any functor \mbox{$\mathfrak{I}:\many\rightarrow\catr$}.
\end{definition}
Recall that this means that $\mathfrak{I}(X)\geq \mathfrak{I}(Y)$ whenever $\mor_{\many}(X,Y)\neq \emptyset.$ Diagrammatically, this is the behavior we seek:
\begin{equation}\label{eq:diagram-invariants}
\xymatrix{
(X,d_X)\ar[r]^{\phi} \ar[d]_{\ifunc} & (Y,d_Y) \ar[d]^{\ifunc} \\
\ifunc(X) \ar[r]^{\geq} & \ifunc(Y)}.
\end{equation}
\subsection{The case of $\miso$.} The collection of invariants in $\miso$ is the largest possible among $\miso,\mmon$ and $\mgen$, and this could be seen as a consequence of the fact that $\miso$ is the category with the fewest morphisms cf. (\ref{eq:inclu-cats}). In this case all $\miso$-functorial invariants are given in the following manner: $\ifunc(X)=\Psi(d_X)$
where $\Psi:\R^{|X|}_+\times\R^{|X|}_+\rightarrow \mathbb{E}$ is such that $\Psi(A) = \Psi(\pi'\cdot A\cdot \pi)$ for any permutation matrix $\pi$ of size $|X|\times |X|$ and $A\in \R^{|X|}_+\times\R^{|X|}_+$.
\subsection{The case of $\mmon$.}
\begin{example} A first example of a $\mmon$-invariant is any non-increasing function $\zeta:\N\rightarrow \mathbb{E}$ of the cardinality $|X|$ of a finite metric space $X$: clearly, since $\mor_{\mmon}(X,Y)\neq \emptyset$ requires that $|X|\leq |Y|$, then $\zeta(|X|)\geq \zeta(|Y|)$. Hence, the construction is functorial.
\end{example}
Another simple example is \mbox{$\ifunc_{\mathrm{sep}}:\mmon\rightarrow \catr$}, which assigns to any finite metric space $X$ its separation $\sep(X).$ This example belongs to a larger class.
\begin{example}
For each $k\geq 2$ let $\ifunc_k:\mmon\rightarrow\catr$ be the functor given by
$$\ifunc_k^-(X):=\inf\big\{\varepsilon\geq 0|\,\mor_{\mmon}(\Delta_k(\varepsilon),X)\neq \emptyset\big\}$$
\mbox{for $X\in\ob(\mmon)$ with $|X|\geq k$}. If $|X|<k$ we put $\ifunc_k(X)=+\infty.$
\end{example}
Notice that $\ifunc_{sep}=\ifunc_2^-$.
That this definition is indeed $\mmon$-functorial can be seen easily.
\begin{proof} Assume that $X,Y\in\ob(\mmon)$ and $\phi\in\mor_{\mmon}(X,Y)$, then, we need to check that $\ifunc_k^-(X)\geq \ifunc_k^-(Y)$. If $|X|,|Y| < k$, or $|X|\geq k$ and $|Y|<k$, there's nothing to prove. Let's assume that $|X|,|Y|\geq k$. Pick $\varepsilon>\ifunc_k^-(X)$ and $f\in\mor_{\mmon}(\Delta_k(\varepsilon),X)$. Then, $\phi\compcirc f \in \mor_{\mmon}(\Delta_k(\varepsilon),Y)$, and hence $\varepsilon\geq \ifunc_k^-(Y).$ The conclusion follows since $\varepsilon>\ifunc_k^-(X)$ was arbitrary.
\end{proof}
Another family of $\mmon$-functorial invariants with a similar structure is the following.
\begin{example}
For each $k\in\N$ define
$$\ifunc_k^+(X):=\sup\big\{\varepsilon\geq 0|\,\mor_{\mmon}(X,\Delta_k(\varepsilon))\neq \emptyset\big\}$$
if $|X|\leq k$, and $\ifunc_k^+(X)=0$ otherwise.
\end{example}
Finally, these two examples can be generalized in the following manner.
\begin{definition}
Let $\Omega\subset \mathcal{M}$ be a collection of finite metric spaces and $\iota:\miso\rightarrow\catr$ any $\miso$-functorial invariant. Then, one defines
$$\mathfrak{I}_\Omega^-(X):=\inf\big\{\iota(\omega)|\,\exists\,\omega\in\Omega\,\mbox{s.t.}\,\mor_{\mmon}(\omega,X)\neq \emptyset\big\}$$
and
$$\mathfrak{I}_\Omega^+(X):=\sup\big\{\iota(\omega)|\,\exists\,\omega\in\Omega\,\mbox{s.t.}\,\mor_{\mmon}(X,\omega)\neq \emptyset\big\}.$$
\end{definition}
In these definitions we use the convention that $\inf$ over an empty set equals $+\infty$ whereas $\sup$ over an empty set equals $0.$ The proof that these these two constructions are $\mmon$-functorial follows the same line of argument as for the case of $\mathfrak{I}_k^-$ above. That these definitions generalize the ones of $\mathfrak{I}^\pm_k$ above can be seen by letting $\Omega=\{\Delta_k(\delta),\,\delta>0\}$ and $\iota$ equal to the separation invariant.
We do not know a complete classification of all the $\mmon$-functorial invariants but it seems that progress could be made using our techniques. Notice that the family of $\mmon$-functorial invariants is rich in the sense that it is closed under addition and multiplication.
\subsection{The case of $\mgen$.} This case is particularly easy to study: since the collection of all morphisms is quite large (it is a super-set of those of $\miso$ and $\mmon$), one would expect that there are relatively few $\mgen$-functorial invariants. This is indeed the case and one has that the only $\mgen$-functorial invariants that are possible are the constant ones: as we saw in Remark \ref{remark:mgen}, one property of the category $\mgen$ is that for any $X,Y\in\ob(\mgen)$ there exist morphisms $\phi_1\in\mor_{\mgen}(X,Y)$ and $\phi_2\in\mor_{\mgen}(Y,X)$. Functoriality of $\mathfrak{I}$ would then require that $\mathfrak{I}(X)\geq \mathfrak{I}(Y)\geq \mathfrak{I}(X),$
and hence, since $X,Y$ were arbitrary, $\mathfrak{I}$ must be constant.
\section{Standard Clustering}
\label{results-sc}
\begin{definition}\label{ex:key} For each $\delta>0$ we define the \textbf{Vietoris-Rips clustering functor}
{$$\rfunc_\delta: \mgen \rightarrow
\catc$$}
as follows. For a finite metric space $(X, d_X)$, we set $\rfunc_\delta(X,d_X)$ to be $(X,P_X(\delta))$, where
$P_X(\delta)$ is the partition of $X$ associated to the equivalence
relation $\sim_\delta$ defined in Example \ref{rips}. We define how $\rfunc_\delta$ acts on maps $f : (X,d_X) \rightarrow (Y,d_Y)$:
$\rfunc_\delta(f)$ is simply the set map $f$ regarded as a
morphism from $(X,P_X(\delta))$ to $(Y,P_Y(\delta))$ in
$\catc$.
\end{definition}
We now check that $\rfunc_\delta$ is in fact $\mgen$-functorial. Fix finite metric spaces $(X,d_X)$ and $(Y,d_Y)$ and a morphism $\phi\in\mor_{\mgen}(X,Y)$. It is enough to prove that whenever $x,x'\in X$ are s.t. $x\sim_\delta x'$, then $\phi(x)\sim_\delta\phi(x')$ in $Y$ as well. Let $x=x_0,x_1,\ldots,x_n=x'$ be points in $X$ with $d_X(x_i,x_{i+1})\leq \delta$ for all $i=0,1,\ldots,n-1$. Then, since $\phi$ is distance non-increasing, $d_Y(\phi(x_i),\phi(x_{i+1}))\leq \delta$ for all $i=0,1,\ldots,n-1$ as well, and by Definition \ref{def-equiv-rips}, then $\phi(x)\sim_\delta\phi(x')$.\footnote{The Vietoris-Rips functor is actually just \textbf{single linkage clustering} as it is well known, see \cite{carlsson-memoli-full,clust-um}. The name comes from certain simplicial constructions used by Vietoris in the 1930s to study \v Cech cohomology theories of metric spaces. Much later, Rips used similar constructions in the context of geometric group theory, see \cite{hausmann}.}
By restricting $\rfunc_\delta$ to the subcategories
$\miso\mbox{ and } \mmon$,
we obtain functors $\rfunc_\delta^{iso}: \miso
\rightarrow \catc$ and $\rfunc_\delta^{inj}:
\mmon \rightarrow\catc$. We will denote all these functors by $\rfunc_\delta$ when there is no ambiguity.
\begin{remark}[\bf The Vietoris-Rips functor is surjective.] Among the desirable conditions singled out by Kleinberg \cite{kleinberg}, one has that of \emph{surjectivity} (which he referred to as ``richness''). Given a finite set $X$ and $P_X\in\mathcal{P}(X)$, surjectivity calls for the existence of a metric $d_X$ on $X$ such that $\rfunc_\delta(X,d_X)=(X,P_X)$.
Indeed this is the case: pick any $\alpha>1$ and let $d_X(x,x')=\delta$ whenever $x$ and $x'$, $x\neq x'$, are in the same block of $P_X$, and $d_X(x,x')=\alpha \delta$ when $x$ and $x'$ are in different blocks of $P_X$. With this definition it is easy to verify that $d_X$ is an ultrametric on $X$ and hence a metric. That $\rfunc_\delta(X,d_X)=(X,P_X)$ follows directly from Lemma \ref{prop:props-uaster} and the definition of the Vietoris-Rips functor.
\end{remark}
For $\many$ being any one of our choices $\miso$,
$\mmon$ or $\mgen$, a clustering functor in this context will be
denoted by $\mathfrak{C}:\many\rightarrow \catc$. \textbf{Excisiveness} of a clustering functor refers to the property that once a finite metric space has been partitioned by the clustering procedure, it should not be further split by subsequent applications of the same algorithm.
\begin{definition}[\bf Excisive clustering functors]
We say that a clustering functor $\cfunc$ is \textbf{excisive} if for all
$(X,d_X) \in \ob(\many)$, if we write $\mathfrak{C}(X,d_X) = (X, \{X_\alpha\}_{\alpha\in
A})$, then
{ $$\cfunc\left(X_\alpha,{d_X}_{|_{X_\alpha \times X_\alpha}}\right) = (X_\alpha,\{X_\alpha\})\,\,\mbox{for all $\alpha\in A$.}$$}
\end{definition}
\begin{remark}[\bf The Vietoris-Rips functor is excisive]\label{remark:rips-excisive}
Fix $\delta>0$ and consider the functor $\rfunc_\delta$ defined in Example \ref{ex:key}. We claim that $\mathfrak{R}_\delta$ is excisive. Indeed, write $\mathfrak{R}_\delta(X,d) = (X,\{X_\alpha,\,\alpha\in A\})$. Then, $X_\alpha$, $\alpha\in A$ are the different blocks of the partition of $X$ into equivalence classes of $\sim_\delta$ ($x\sim_\delta x'$ if and only if $x,x'\in X_\alpha$ for some $\alpha\in A$). Fix $\alpha\in A$ and let $x,x'\in X_\alpha$. Let $x_0,x_1,\ldots,x_m\in X_\alpha$ be s.t. $x_0=x$, $x_m=x'$ and $d_X(x_i,x_{i+1})\leq \delta$ for $i\in\{0,\ldots,m-1\}$. But then it follows that $x_i\in X_\alpha$ for all $i\in\{0,\dots,m\}$ as well and hence $\mathfrak{R}_\delta\big(X_\alpha,{d_X}_{|_{X_\alpha\times X_\alpha}}\big)=(X_\alpha,\{X_\alpha\}).$
\end{remark}
There exist large families of non-excisive clustering functors in $\mmon$.
\begin{example}[\bf A family of non-excisive functors in $\mmon$]\label{example:non-excisive}
For each invariant $\ifunc:\mmon\rightarrow\catr$ (recall their construction in \S\ref{sec:invariants}) and non-increasing function $\eta:\mathbb{E}\rightarrow\mathbb{E}$, consider the clustering functor $\widehat{\rfunc}:\mmon\rightarrow \catc$ defined as follows: For a finite metric space $(X, d_X)$, we define $\widehat{\rfunc}(X, d_X)$ to be $(X,\widehat{P}_X)$,
where $\widehat{P}_X$ is the partition of $X$ associated to the equivalence relation $\sim_{\eta(\ifunc(X))}$ on $X$. That $\widehat{\rfunc}$ is a functor follows from the fact that whenever $\phi\in\mor_{\mmon}(X,Y)$ and $x\sim_{\eta(\ifunc(X))} x'$, then $\phi(x)\sim_{\eta(\ifunc(Y))}\phi(x')$.
Indeed, let $x,x'\in X$ be s.t. $x\sim_{\eta(\ifunc(X))}x'$ and let $x=x_0,\ldots,x_n=x'$ in $X$ s.t. $d_X(x_i,x_{i+1})\leq \eta(\ifunc(X))$ for $i=0,\dots,n-1$. Then, for $i=0,\ldots,n-1$,
$d_Y(\phi(x_i),\phi(x_{i+1}))\leq d_X(x_i,x_{i+1})\leq \eta(\ifunc(X)).$
Since $\eta$ is non-increasing and $\ifunc(X)\geq \ifunc(Y)$, it follows that for $i=0,1,\ldots,n-1$,
$d_Y(\phi(x_i),\phi(x_{i+1}))\leq \eta(\ifunc(Y))$,
and hence $\phi(x)\sim_{\eta(\ifunc(Y))}\phi(x').$
Now, the functors $\widehat{\rfunc}$ are \textbf{not excisive} in general. An explicit example is the following, which applies to $\ifunc=\ifunc_{sep}$ and $\eta(z)=z^{-1}$. Consider the metric space $(X,d_X)$ depicted in figure on the right, where the metric is given by the graph metric on the underlying graph (the edge weights are shown in the figure). Note that $\textrm{sep}(X,d_X)=1/2$ and thus $\eta(\ifunc(X)) = 2.$ We then find that $\widehat{\mathfrak{R}}(X,d_X) = (X,\{\{A,B,C\},\{D,E\}\})$. But,
$\sep\left(\{A,B,C\},\dmthree{2}{3}{1} \right) = 1$
and hence $\eta(\ifunc({\{A,B,C\}}))=1$. Therefore,
{ $$\widehat{\rfunc}\left(\{A,B,C\},\dmthree{2}{3}{1} \right)=(\{A,B,C\},\{A,\{B,C\}\}),$$}
and we see that $\{A,B,C\}$ gets further partitioned by $\widehat{\rfunc}$.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{non-excisive.eps}
\caption{Metric space used to prove that the functor $\widehat{\rfunc}:\mmon\rightarrow \catc$ is not excisive. The metric is given by the graph distance on the graph.} \label{fig:non-excisive}
\end{figure}
\end{example}
It is interesting to point out that the same construction of a non-excisive functor in $\mgen$ would not work, since as we saw in \S\ref{sec:invariants}, all the invariants $\mathfrak{I}:\mgen\rightarrow\catr$ are constant. We will see in \S\ref{sec:mgen} that $\mgen$ permits the existence of only one (well behaved) clustering functor which turns out to be the Vietoris-Rips functor, which is excisive as we saw in Remark \ref{remark:rips-excisive}.
\subsection{The case of \texorpdfstring{$\miso$}{Miso}}\label{sec:miso}
It is easy to describe all $\miso$-functorial clustering schemes. Let $\mathcal{I}$ denote the collection of all isometry classes of finite metric spaces. Clearly, one must specify the clustering functor on each class of $\mathcal{I}$ in a manner which is invariant to self-symmetries of (an element of) the class. The precise statement is as follows. For each $\zeta\in\mathcal{I}$ let $(X_\zeta,d_{X_\zeta})$ denote an element of the class $\zeta$, $G_\zeta$ the isometry group of $(X_\zeta,d_{X_\zeta})$, and $\Xi_\zeta$ the set of all fixed points of the action of $G_\zeta$ on $\mathcal{P}(X_\zeta)$. For completeness we state as a theorem the following obvious fact:
\begin{theorem}[\textbf{Classification of $\miso$-functorial clustering schemes}]\label{theo:miso}
Any $\miso$-functorial clustering scheme determines a choice of $p_\zeta\in\Xi_\zeta$ for each $\zeta\in\mathcal{I}$, and conversely, a choice of $p_\zeta$ for each $\zeta\in\mathcal{I}$ determines an $\miso$-functorial scheme.
\end{theorem}
\subsection{Representable Clustering Functors}
In what follows, $\many$ is either of $\mmon$ or $\mgen$. For each $\delta>0$ the Vietoris-Rips functor $\rfunc_\delta:\many\rightarrow\catc$ can be described in an alternative way which relies on the ability to construct maps from the two point metric space $\Delta_2(\delta)$ into a given input finite metric space $(X,d_X)$. A first trivial observation is that the condition that $x,x'\in X$ satisfy $d_X(x,x')\leq \delta$ is equivalent to requiring the existence of a map $f\in{\mor}_{\many}(\Delta_2(\delta),X)$ with $\{x,x'\}\subset \im(f)$. Using this observation, we can reformulate the condition that $x\sim_\delta x'$ by the requirement that there exist $z_0,z_1,\ldots,z_k\in X$ with $z_0=x$, $z_k=x'$, and $f_1,f_2,\ldots,f_k\in\mor_{\many}(\Delta_2(\delta),X)$ with $\{x_{i-1},x_i\}\subset\im(f_i)$ $\forall i=1,2,\ldots,k$. Informally, this points to the interpretation that $\{\Delta_2(\delta)\}$ is the ``parameter'' in a ``generative model'' for $\rfunc_\delta$.
This immediately suggests considering more general clustering functors constructed in the following manner. Let $\Omega$ be any fixed collection of finite metric spaces. Define a clustering functor \mbox{$\mathfrak{C}^\Omega:\many\rightarrow\catc$} as follows: let
$(X,d)\in \ob(\many)$ and write $\mathfrak{C}^\Omega(X,d) = (X,\{X_\alpha\}_{\alpha\in
A})$. One declares that points $x$ and $x'$ belong to the same
block $X_\alpha$ if and only if there exist
\begin{itemize}
\item a sequence of points $z_0,\ldots,z_k\in X$ with $z_0 = x$ and $z_k=x'$,
\item a sequence of metric spaces $\omega_1,\ldots,\omega_k\in
\Omega$ and
\item for each $i=1,\ldots,k$, pairs of points
$(\alpha_i,\beta_i)\in \omega_i$ and morphisms $f_i\in
{\mor}_{\many}(w_i,X)$ s.t. $f_i(\alpha_i) = z_{i-1}$ and
$f_i(\beta_i) = z_i$.
\end{itemize}
Also, we declare that $\cfunc^\Omega(f)=f$ on morphisms $f$. Notice that above one can assume that $z_0,z_1,\ldots,z_k$ all belong to $X_\alpha$.
\begin{remark}\label{remark:rep-no-split}
Let $\Omega$ be a collection of finite metric spaces. Consider the clustering functor \mbox{$\mathfrak{C}^\Omega:\many\rightarrow \catc$} constructed as above. Then, clearly, $\mathfrak{C}^\Omega(\omega,d_\omega)=(\omega,\{\omega\})$ for all $\omega\in\Omega$.
As a consequence, we have the following: If $X$ is any finite metric space and $\omega\in\Omega$, then all $f\in \mor_{\many}(\omega,X)$ are s.t. $\im(f)$ is fully contained in a single block of the partition of $X$ by $\mathfrak{C}^\Omega$.
\end{remark}
\begin{remark}\label{remark:omega}
Let $\Omega\supset \Omega'\neq \emptyset$, then $\mathfrak{C}^{\Omega'}$ produces partitions that are not coarser than those produced by $\mathfrak{C}^{\Omega}$.
\end{remark}
\begin{definition}
We say that a clustering functor $\mathfrak{C}$ is \textbf{representable} whenever
there exists a collection of finite metric spaces $\Omega$
such that $\mathfrak{C} = \mathfrak{C}^\Omega$. In this case, we say that $\mathfrak{C}$ is
\textbf{represented} by $\Omega$. We say that $\cfunc$ is \textbf{finitely representable} whenever $\cfunc=\cfunc^\Omega$ for some \emph{finite} collection of finite metric spaces $\Omega$.
\end{definition}
As we saw above, the Vietoris-Rips functor $\rfunc_\delta$ is (finitely) represented by $\{\Delta_2(\delta)\}$.
\subsection{Representability and Excisiveness.} We now prove that representability and excisiveness are equivalent properties. Notice that excisiveness is an axiomatic statement whereas representability asserts existence of generative model for the clustering functor.
\begin{theorem} \label{theo:exc-rep}
Let $\many$ be either of $\mmon$ or $\mgen$. Then any
clustering functor on $\many$ is excisive if and only if it is
representable.
\end{theorem}
\begin{proof}
Assume that $\mathfrak{C}$ is an excisive clustering functor. Let
\beq{eq:omega}\Omega:=\big\{(B,d_{|_{B\times B}}),\, B\in P,\, \mbox{where $(X,P)=\mathfrak{C}(X,d)$ and $(X,d)\in \mathcal{M}$}\big\}.\eeq
That is, $\Omega$ contains all the possible blocks (endowed with restricted metrics) of all partitions obtained by applying $\mathfrak{C}$ to all finite metric spaces. Notice that by definition of $\Omega$, and since $\mathfrak{C}$ is excisive,
\beq{eq:excisive-omega}\mathfrak{C}(\omega,d_\omega) = (\omega,\{\omega\})\,\mbox{for all $\omega\in\Omega$}.\eeq
We will now prove that $\mathfrak{C}$ is represented by $\Omega$, that is $\mathfrak{C}=\mathfrak{C}^\Omega$. Fix any finite metric space $(X,d_X)$ and write $(X,P)=\mathfrak{C}(X,d_X)$ and $(X,P')=\mathfrak{C}^\Omega(X,d_X)$. We know that $\Omega$ contains all blocks $B\in P$, and since the inclusion $B\hookrightarrow X$ is a morphism in $\mmon$, by Remark \ref{remark:rep-no-split} one sees that $\mathfrak{C}^\Omega(B,{d_{X}}_{|_{B\times B}})=(B,\{B\})$. Now, by functoriality of $\mathfrak{C}^\Omega$, any two points $x,x'$ in $B$ must belong to the same block of $P'$. Since $B$ was any block of $P$ it follows that $P$ is a refinement of $P'$.
This is depicted by the diagram below:
\begin{equation}\label{eq:diagram-i}
\xymatrix{
(B,{d_{X}}_{|_{B\times B}}) \ar[r]^{\hookrightarrow} \ar[d]_{\mathfrak{C}^\Omega} & (X,d_X) \ar[d]^{\mathfrak{C}^\Omega} \\
(B,\{B\}) \ar[r]^{\mathfrak{C}^\Omega(\hookrightarrow)} & (X,P')}
\end{equation}
Conversely, assume $x,x'\in B'\in P'$ and let $\omega_1,\ldots,\omega_k\in \Omega$, $(\alpha_i,\beta_i)\in \omega_i$ for $i=1,2,\ldots,k$, $z_0,z_1,\ldots,z_k \in X$ and $f_1,f_2,\ldots,f_k$ s.t. $f_i\in \mor_{\many}(\omega_i,X)$, $f_i(\alpha_i)=z_{i-1}$ and $f_i(\beta_i)=z_i$ for $i=1,2,\ldots,k$. Consider for each $i\in\{1,2,\ldots,k\}$ the diagram on the right, where we have used (\ref{eq:excisive-omega}). Then, since for each $i\in\{1,\ldots,k\}$ $\alpha_i,\beta_i$ are in the same (unique) block of $\omega_i$, $z_{i-1}$ and $z_i$ have to be in the same block of $P$.
$$\xymatrix{
(\omega_i,d_{\omega_i}) \ar[r]^{f_i} \ar[d]_{\mathfrak{C}} & (X,d_X) \ar[d]^{\mathfrak{C}} \\
(\omega_i,\{\omega_i\}) \ar[r]^{\mathfrak{C}(f_i)} & (X,P)}.$$
Since this has to happen for all $i$, we conclude that $x=z_0$ and $z_k=x'$ ought to be in the same block of $P$. Thus $P'$ is a refinement of $P$. We have proved that $P$ refines and is refined by $P'$ and therefore these two partitions must be equal. Since $X$ was any finite metric space, we conclude that $\mathfrak{C}$ is represented by $\Omega$ and we are done.
Assume now that $\cfunc$ is representable, i.e. $\mathfrak{C}=\mathfrak{C}^\Omega$ for some collection of finite metric spaces $\Omega$. We will prove that $\cfunc$ is excisive. Fix a finite metric space $(X,d_X)$ and write $\mathfrak{C}^\Omega(X,d_X)=(X,P)$. Pick any block $B$ of $P$ and any two points $x,x'\in B$. Let $\omega_1,\ldots,\omega_k\in \Omega$,
$f_1\in\mor_{\many}(\omega_1,X),\ldots,f_k\in\mor_{\many}(\omega_k,X);$ $(\alpha_1,\beta_1)\in\omega_1,\ldots,(\alpha_k,\beta_k)\in\omega_k$; and $z_0,z_1,\ldots,z_k$ be s.t. $z_0=x$, $z_k=x'$, $f_i(\alpha_i)=z_{i-1}$, $f_i(\beta_i)=z_i$ for all $i=1,2,\ldots,k$. One can assume that all $z_i\in B$ and moreover, by Remark \ref{remark:rep-no-split}, $\im(f_i)\subseteq B$. Hence, we can regard each $f_i$ as a morphism in $\mor_{\many}(\omega_i,B)$ where $B$ is endowed with the restricted metric.
By Remark \ref{remark:rep-no-split}, $\cfunc(\omega,d_\omega)=(\omega,\{\omega\})$ for all $\omega\in\Omega$. Hence, by functoriality we have that $x,x'\in B$ are in the same block of $P_B$, where $(B,P_B) = \mathfrak{C}^\Omega(B,{d_X}_{|_{B\times B}})$. Since $x,x'\in B$ were arbitrary we conclude that $P_B$ consists of only one block. The arbitrariness in the choices of $B\in P$ and $X\in\ob(\many)$ finishes the proof that $\cfunc=\mathfrak{C}^\Omega$ is excisive.
\end{proof}
\subsection{A factorization theorem.} \label{section:factor}
We now prove that all finitely representable clustering functors in $\mgen$ and $\mmon$ arise as the composition of the Vietoris-Rips functor with a functor that changes the metric.
For a given collection $\Omega$ of finite metric spaces let
\beq{eq:Tfunc}\tfunc^\Omega:\many\rightarrow \many
\eeq be the endofunctor that assigns to each finite metric space $(X,d_X)$ the metric space $(X,d_X^\Omega)$ with the same underlying set and metric given by $d_X^\Omega=\mathcal{U}(W_X^\Omega)$,\footnote{Recall (\ref{eq:U}), the definition of $\mathcal{U}$.} where $W_X^\Omega:X\times X\rightarrow \R_+$ is given by
\beq{eq:W}(x,x')\mapsto \inf\big\{\lambda>0|\,\exists w\in\Omega\,\mbox{and}\,\phi\in\mor_{\many}(\lambda\cdot\omega,X)\,\mbox{with $\{x,x'\}\subset\im(\phi)$}\big\},\eeq
for $x\neq x'$, and by $0$ on $\mathrm{diag}(X\times X)$. Above we assume that the $\inf$ over the empty set equals $+\infty$. Note that $W_X^\Omega(x,x')<\infty$ for all $x,x'\in X$ as long as $|\omega|\leq |X|$ for some $\omega\in \Omega$. Also, $W_X^\Omega(x,x')=\infty$ for all $x\neq x'$ when $|X|<\inf\{|\omega|,\,\omega\in\Omega\}.$
\begin{theorem}\label{theo:comp-rips}
Let $\many$ be either $\mgen$ or $\mmon$ and $\cfunc$ be any $\many$-functorial finitely representable clustering functor represented by some $\Omega\subset \mathcal{M}$. Then, $\cfunc=\rfunc_1\compcirc\tfunc^\Omega$.
\end{theorem}
\begin{proof}
Fix any finite metric space $(X,d_X)$ and write $\cfunc(X,d_X)=(X,P)$ and $\rfunc_1\compcirc\tfunc^\Omega(X,d_X)=(X,P').$
Assume that $x,x'\in B\in P$ and let $x=z_0,z_1,\ldots,z_k=x'\in X$, $\omega_1,\ldots,\omega_k\in \Omega$, $f_i\in\mor_{\many}(\omega_i,X)$ and $(\alpha_i,\beta_i)\in\omega_i$ s.t. $f_i(\alpha_i)=z_{i-1}$, $f_i(\beta_i)=z_i$ for $i=1,2,\ldots,k.$ Then, from (\ref{eq:W}) for all $i=1,2,\ldots,k$, $W_X^\Omega(z_{i-1},z_{i})\leq 1$ and in particular $\max_i W_X^\Omega(z_{i-1},z_i)\leq 1$. It follows from (\ref{eq:U}) that $d_X^\Omega(x,x')\leq 1$ and consequently $x$ and $x'$ are in the same block of $P'$. Thus we see that $P$ refines $P'$.
Assume now that $x,x'\in B'\in P'$. By definition of the Vietoris-Rips functor, $x\sim_1 x'$ in $(X,d_X^\Omega)$, which implies the existence of $x=x_0,x_1,\ldots,x_k=x'$ in $X$ with $d_X^\Omega(x_i,x_{i+1})\leq 1$ for all $i=1,\ldots,k-1$. Now, by (\ref{eq:U}), for each $i=0,1,\ldots,k$ there exist $y_0^{\scriptscriptstyle(i)},y_1^{\scriptscriptstyle (i)},\ldots,y_{k_i}^{\scriptscriptstyle(i)}$ in $X$ with $y_0^{\scriptscriptstyle(i)}=x_i$, $y_{k_i}^{\scriptscriptstyle(i)}=x_{i+1}$ and $W_X^\Omega\big(y_{j}^{\scriptscriptstyle(i)},y_{j+1}^{\scriptscriptstyle(i)}\big)\leq 1$ for $j=0,1,\ldots,k_i-1$. By concatenating all sequences $\big\{y_j^{\scriptscriptstyle(i)}\big\}_{j=1}^{k_i}$ for $i=0,1,\ldots,k$, we obtain a sequence $z_0,z_1,\ldots, z_N \in X$ with $z_0=x$, $z_N=x'$ and $W_X^\Omega(z_i,z_{i+1})\leq 1$. By definition of $W_X^\Omega$ and because $\Omega$ is finite, one sees that for every $i=1,2,\ldots, N$ there exists $\lambda_i\in(0,1]$, $\omega_i\in\Omega$ and $\phi_i\in\mor_{\many}(\lambda_i\cdot\omega_i,X)$ with $\{z_{i-1},z_i\}\subset\im(\phi_i).$ But then, obviously, $\phi_i\in\mor_{\many}(\omega_i,X)$ as well, and since $\cfunc$ is represented by $\Omega$ one sees that $x$ and $x'$ must be in the same block of $P.$ This proves that $P'$ refines $P$, and finishes the proof.
\end{proof}
\subsection{The case of \texorpdfstring{$\mgen$}{Mgen}}\label{sec:mgen}
\subsubsection{A Uniqueness theorem.} In $\mgen$ clustering functors are very restricted, as reflected by the following theorem.
\begin{theorem}\label{theo:unique}
Assume that $\mathfrak{C}:\mgen\rightarrow\catc$ is a clustering functor for which there exists $\delta_{\mathfrak{C}}> 0$ with the property that
\begin{itemize}
\item $\mathfrak{C}(\Delta_2(\delta))$ is in one piece for all $\delta\in [0,\delta_\mathfrak{C}]$, and
\item $\mathfrak{C}(\Delta_2(\delta))$ is in two pieces for all $\delta> \delta_\mathfrak{C}$.
\end{itemize}
Then, $\cfunc$ is the Vietoris-Rips functor with parameter $\delta_{\mathfrak{C}}$. i.e. $\cfunc=\rfunc_{\delta_\cfunc}$.
\end{theorem}
Recall that the Vietoris-Rips functor is excisive, cf. Remark \ref{remark:rips-excisive}.
\begin{remark}[About uniqueness in $\mgen$]\label{app:rips-strict}
For each $\delta>0$ consider the clustering functor $\stackrel{\circ}{\rfunc}_\delta$ that assigns points $x,x'$ in a metric space $(X,d_X)$ to the same block if and only if there exist $x_0,x_1,\ldots,x_k\in X$ with $x_0=x$, $x_k=x'$ and $d_X(x_i,x_{i+1})<\delta$ for $i=0,1,\ldots,k-1$. Note the \emph{strict} inequality, cf. with the definition of the Vietoris-Rips functor in Example \ref{ex:key}.
If one changes the conditions in Theorem \ref{theo:unique} in a way such that
\begin{itemize}
\item $\mathfrak{C}(\Delta_2(\delta))$ is in one piece for all $\delta\in [0,\delta_\mathfrak{C})$, and
\item $\mathfrak{C}(\Delta_2(\delta))$ is in two pieces for all $\delta\geq \delta_\mathfrak{C}$,
\end{itemize}
then, $\cfunc=\stackrel{\circ}{\rfunc}_{\delta_\cfunc}$.
\end{remark}
\begin{proof}[Proof of Theorem \ref{theo:unique}]
Fix a finite metric space $(X,d_X)$ and write $(X,P)=\cfunc(X,d_X)$ and $(X,P')=\rfunc_{\delta_\cfunc}(X,d_X).$
\noindent
(1) Let $x,x'\in B$ for some block $B\in P'$. We will prove that then $x,x'$ are contained in the same block of $P$. Let $x_0,x_1,\ldots,x_k\in X$ be s.t. $x_0=x$, $x_k=x'$ and $d_X(x_i,x_{i+1})\leq \delta_\cfunc$ for all $i\in\{1,\ldots,k-1\}$. We now prove that each pair $\{x_i,x_{i+1}\}$ belongs to the same block of $P'$, what will yield the claim. Indeed, since for each $i$, $d_X(x_i,x_{i+1})\leq \delta_\cfunc$, then for each $i$ there exists $f_i\in\mor_{\mgen}(\Delta_2(\delta_\cfunc),X)$ with $\im(f_i) = \{x_i,x_{i+1}\}$. Thus, by functoriality and the definition of $\delta_\cfunc$, $x_i$ and $x_{i+1}$ must fall in the same block of $P'$. This concludes the first half of the proof.
\noindent
(2) Assume that $x_0,x_0'\in X$ are in different blocks of $P'$. We will prove that then $x_0$ and $x_0'$ must belong to different blocks of $P$.
Let $\delta := \min\big\{d_X(x,x'),\,x\in X\backslash[x_0']_{\delta_\cfunc},\,x'\in[x_0']_{\delta_\cfunc}\big\}.$
Notice that by Lemma \ref{prop:props-uaster}, $\delta>\delta_\cfunc$. Write
$\Delta_2(\delta) = \Big(\{p,q\},\dmtwo{\delta}\Big)$
and define the map \mbox{$\phi:X\rightarrow \Delta_2(\delta)$} given by $\phi(x) = p$ for all $x\in X\backslash [x_0']_{\delta_\cfunc}$, and $\phi(x') = q$ for all $x'\in [x_0']_{\delta_\cfunc}$. We claim that $\phi\in\mor_{\mgen}(X,\Delta_2(\delta))$. Indeed, if $x,x'\in X\backslash [x_0']_{\delta_\cfunc}$ or $x,x'\in [x_0']_{\delta_\cfunc}$, $\phi(x)=\phi(x')$ and there's nothing to check. Assume then that $x\in X\backslash [x_0']_{\delta_\cfunc}$ and $x'\in [x_0']_{\delta_\cfunc}$, and note that by definition of $\delta$,
$d_X(x,x')\geq\delta$. On the other hand, $p=\phi(x)$ and $q=\phi(x')$ are at distance $\delta$ from each other. Thus $\phi$ is distance non-increasing. Now, we apply $\cfunc$ to $X$ and $\Delta_2(\delta)$ and note that in this case, by our assumption on $\cfunc$, since $\delta>\delta_\cfunc$, then $\cfunc(\Delta_2(\delta))$ is in two pieces. By functoriality we conclude that $x_0$ and $x_0'$ cannot lie in the same block of $P.$
\end{proof}
\subsection{Scale invariance in $\mgen$ and $\mmon$.} It is interesting to consider the effect of imposing Kleinberg's scale invariance axiom on $\mgen$-functorial and $\mmon$-functorial clustering schemes. It turns out that in $\mgen$ there are only two possible clustering schemes enjoying scale invariance, which turn out to be the trivial ones:
\begin{theorem}\label{theorem:scale-invariance}
Let $\cfunc:\mgen\rightarrow\catc$ be a clustering functor s.t. $\cfunc\compcirc \sigma_\lambda=\cfunc$ for all $\lambda>0$. Then, either
\begin{itemize}
\item $\cfunc$ assigns to each finite metric space $X$ the partition of $X$ into singletons, or
\item $\cfunc$ assigns to each finite metric the partition with only one block.
\end{itemize}
\end{theorem}
\begin{proof}Let $\cfunc:\mgen\rightarrow\catc$ be a scale invariant clustering functor. Consider first the metric space $\Delta_2(1)$ consisting of two points at distance $1$. There are two possiblities: either \textbf{(1)} $\cfunc(\Delta_2(1))$ is in one piece, or \textbf{(2)} $\cfunc(\Delta_2(1))$ in two pieces.
Assume \textbf{(1)}. By scale invariance, the partition of $\Delta_2(\delta)$ given by $\cfunc(\Delta_2(\delta))$ is in one piece for all $\delta>0$. Fix any finite metric space $(X,d_X)$ and write $\cfunc(X,d_X) = (X,P)$. Pick any $x,x'\in X$, let $\delta = d_X(x,x')$, and consider the morphism $\phi\in\mor_{\mgen}(\Delta_2(\delta),X)$ s.t. $\phi(p)=x$ and $\phi(q)=x'$, where $p$ and $q$ are the two elements of $\Delta_2(\delta).$ By scale invariance and functoriality, $x$ and $x'$ must then be in the same block of $P$. Since $x$ and $x'$ were arbitrary, $P$ must be in one piece as well. It follows from the arbitrariness of $X$ that $\cfunc$ then assigns the single block partition to any finite metric space.
Now assume \textbf{(2)}. By scale invariance $\cfunc(\Delta_2(\delta))$ is in two pieces for all $\delta>0$. Pick any finite metric space $(X,d_X)$ and write $\cfunc(X,d_X) = (X,P)$. Pick any $x,x'\in X$, and a partition of $X$ into subsets $A$
and $B$ such that $x\in A$ and $x'\in B$. Let $\delta =\sep(X)$, and consider the map $\psi:X\rightarrow \Delta_2(\delta)$ given by $\psi(a)=p$ for all $a\in A$ and $\psi(b)=q$ for all $b\in B$. It is easy to verify that $\psi\in\mor_{\mgen}(X,\Delta_2(\delta))$. By scale invariance and functoriality, $x$ and $x'$ cannot be in the same block of $P$. Since $x$ and $x'$ were arbitrary, no two points belong to the same block of $P$, and hence $\cfunc$ assigns $X$ the partition into singletons. It follows from the arbitrariness of $X$ that $\cfunc$ assigns to every finite metric space its partition into singletons. This finishes the proof.
\end{proof}
By refining the proof of the previous theorem, we find that the behavior of any $\mmon$-functorial clustering functor is also severely restricted in each $\mathcal{M}_k$, $k\in\N$.\footnote{Recall that $\mathcal{M}_k$ denotes the collection of all $k$-point metric spaces.}
\begin{theorem}\label{theorem:scale-invariance-mmon}
Let $\cfunc:\mmon\rightarrow\catc$ be a clustering functor s.t. $\cfunc\compcirc \sigma_\lambda=\cfunc$ for all $\lambda>0$. Let $$K(\cfunc):=\big\{k\geq 2;\,\cfunc(\Delta_k(1))\,\mbox{is in one piece}\big\}.$$
\begin{itemize}
\item If $K(\cfunc)=\emptyset$, then $\cfunc$ assigns to each finite metric space $X$ its partition into singletons.
\item Otherwise, let $k_\cfunc=\min K(\cfunc)$. Then,
\begin{itemize}
\item For each $k\geq k_\cfunc$, $\cfunc$ assigns to each finite metric space $X\in \mathcal{M}_k$ the partition of $X$ with only one block.
\item For each $2\leq k< k_\cfunc$, $\cfunc$ assigns to each finite metric space $X\in \mathcal{M}_k$ the partition of $X$ into $k$ singletons.
\end{itemize}
\end{itemize}
\end{theorem}
\begin{proof}
Assume first that $K(\cfunc)=\emptyset$, pick any finite metric space $X$ and let $k=|X|$. Pick any two distinct points $x,x'\in X$. Choose any injective map $\phi:X\rightarrow \Delta_k(\sep(X))$ such that $\phi(x)$ and $\phi(x')$ fall in different blocks of the partition of $\Delta_k(\sep(X))$ produced by $\cfunc$. Then, functoriality guarantees that $x$ and $x'$ must belong to different blocks of the partition of $X$ generated by $\cfunc$. Since $x,x'\in X$ were arbitrary we conclude that $\cfunc$ partitions $X$ into singletons.
Now assume that $K(\cfunc)\neq\emptyset$. Pick $k\geq k_\cfunc$ and $X\in\mathcal{M}_k$. Consider any morphism $\phi\in\mor_{\mmon}(\Delta_k(\diam(X)),X)$ arising from arbitrarily choosing an injection from the underlying set of $\Delta_k(1)$ to $X$. Then, by functoriality, this forces $\cfunc(X,d_X)$ to produce a partition of $X$ into exactly one block.
Finally, for $2\leq k<k_\cfunc$, $\cfunc$ partitions $\Delta_k(1)$ into more than piece. An argument similar to the one given for the case $K(\cfunc)=\emptyset$ proves that then $\cfunc$ partitions any $X\in\mathcal{M}_k$ into ($k$) singletons.
\end{proof}
\subsection{Clustering functors on \texorpdfstring{$\mmon$ and density}{Mmon}}\label{section:mmon-density}
In the case of $\mmon$, there is a richer class of
allowed clustering schemes that includes functors other than the Vietoris-Rips one. We already saw in Example \ref{example:non-excisive} that in $\mmon$ there are non-excisive clustering schemes that by construction are different from the Vietoris-Rips functor. A class of excisive $\mmon$-functorial clustering schemes which has practical interest is the following.
\begin{figure}[htb] \centering
\includegraphics[width=1\linewidth]{example-rep_s.eps}
\caption{Fix $\delta>0$ and consider the two clustering functors $\mathfrak{C}^{\Delta_2(\delta)}$ and $\mathfrak{C}^{\Delta_3(\delta)}$ on $\mmon$. Then $\mathfrak{C}^{\Delta_2(\delta)}$ applied to both metric spaces in the Figure will yield a single cluster. Meanwhile, $\mathfrak{C}^{\Delta_3(\delta)}$ will yield a single cluster for the metric space on right, but it will not connect the points of the metric space on the left. This admits the interpretation that in order for $\cfunc^{\Delta_3(\delta)}$ to produce a cluster, a certain degree of agglomeration of the points is necessary. This reflects the fact that this clustering scheme is more sensitive to density than its two-point counterpart $\cfunc^{\Delta_2(\delta)}=\rfunc_\delta$. Similar considerations apply to the general case of $\cfunc^{\Delta_k(\delta)}$, any $k\in\N$} \label{fig:ex-rep} \end{figure}
Consider for each $n\in\mathbb{N}$ and
$\delta\in \mathbb{R}_+$ the clustering functor
$\mathfrak{C}^{\Delta_{n}(\delta)}$, and notice that this clustering method implements the notion that
in order to deem a two points $x$ and $x'$ in a finite metric space $X$ to be in the same cluster one requires
that they are ``tightly packed together'' in the sense that there exists a sequence of points $x_0,x_1,\ldots,x_k$ connecting $x$ to $x'$ such that each consecutive pair $\{x_i,x_{i+1}\}$ is contained in a subset $S_i$ of $X$ of cardinality $n$ whose inter-point distances do not exceed $\delta$. This notion is closely related to proposals in network clustering \cite{hungarian-clique} and the DBSCAN algorithm \cite{dbscan}, and for $n>2$ provides a way to tackle the so called ``chaining effect'' exhibited by single linkage clustering (i.e. $\cfunc^{\Delta_2(\delta)}$), see Figure \ref{fig:ex-rep}. Note that Theorem \ref{theo:comp-rips} implies that for a given finite metric space $(X,d_X)$, $\cfunc^{\Delta_3(\delta)}(X,d_X)$ can be obtained by first applying a change of metric to $X$ and then applying the standard Vietoris-Rips functor to the resulting metric space. The change of metric specified by the theorem leads to finding all $3$-cliques in $X$ whose diameter is not greater than $\delta$.
\begin{corollary}\label{coro:change-metric}
On $\mmon$, for all $m\in \N$
$$\cfunc^{\Delta_m(\delta)} = \rfunc_\delta\compcirc \tfunc^m.$$
\end{corollary}
Above, we have used the shorthand notation $\tfunc^m = \tfunc^\Omega$, for $\Omega=\{\Delta_m(\delta),\,\delta\geq 0\}$.
The formalism of representable clustering functors allows for far more generality than that provided by families $\{\Delta_n(\delta),\,n\in\N,\,\delta>0\}$. For example, one could encode the requirement that clusters be densely connected by considering a clustering functor represented by the metric space in Figure \ref{fig:subdiv}t, which consists of the vertices of an equilateral triangle of side $\delta$, together with its center.
\begin{figure}
\centering
\includegraphics[width=.5\linewidth]{subdiv.eps}
\caption{A metric space that encodes a notion of ``density'' different from that provided by $\Delta_3(\delta)$. \label{fig:subdiv}. }
\end{figure}
\section{Hierarchical Clustering}
\label{results-hc}
\begin{example}[\textbf{A hierarchical version of the Vietoris-Rips functor}] \label{ex:key} We define a functor
$$\rfunc: \underline{{\mathcal M}}^{gen} \rightarrow
\underline{{\mathcal P}}$$
as follows. For a finite metric space $(X, d_X)$, we define $(X,d_X)$ to be the persistent set $(X, \theta_X ^{{ \mathrm{VR}}})$, where
$\theta_X^{{ \mathrm{VR}}}(r)$ is the partition associated to the equivalence
relation $\sim _r$ defined in Example \ref{rips}. This is clearly an
object in $\underline{{\mathcal P}}$. We also define how $\rfunc$ acts on maps $f : (X,d_X) \rightarrow (Y,d_Y)$: The value of
$\rfunc(f)$ is simply the set map $f$ regarded as a
morphism from $(X,\theta_X^{{ \mathrm{VR}}})$ to $(Y,\theta_Y^{{ \mathrm{VR}}})$ in
$\underline{{\mathcal P}}$. That it is a morphism in $\underline{{\mathcal
P}}$ is easy to check.
Clearly, this functor implements the hierarchical version of single linkage clustering in the sense that for each $\delta\geq 0$, if one writes $\rfunc_\delta{(X,d_X)}=(X,P_X(\delta))$, then $P_X(\delta) = \theta_X^{{ \mathrm{VR}}}(\delta)$.
\end{example}
\subsection{Functoriality over $\miso$}
This is the smallest category we will deal with. The morphisms in
$\underline{{\mathcal M}}^{iso}$ are simply the bijective maps between
datasets which preserve the distance function. As such, functoriality
of a clustering algorithms over $\underline{{\mathcal M}}^{iso}$ simply means
that the output of the scheme doesn't depend on any artifacts in the
dataset, such as the way the points are named or the way in which they
are ordered.
{{ Agglomerative hierarchical clustering}, in standard
form, as described for example in \cite{clusteringref}, begins
with point cloud data and constructs a \emph{binary} tree (or
dendrogram) which describes the merging of clusters as a
threshold is increased. The lack of functoriality comes from
the fact that when a single threshold value corresponds to
more than one data point, one is forced to choose an ordering
in order to decide which points to ``agglomerate'' first.
This can easily be modified by relaxing the requirement that
the tree be binary. This is what we did in Example
\ref{hierarchical} In this case, one can view these methods as
functorial on $\underline{{\mathcal M}}^{iso}$, where the functor
takes its values in arbitrary rooted trees. It is understood
that in this case, the notion of morphism for the output
($\underline{\mathcal P}$) is simply isomorphism of rooted
trees. In contrast, we see next that amongst these methods,
when we impose that they be functorial over the larger (more
demanding) category $\underline{\mathcal M}^{gen}$ then only
one of them passes the test.\footnote{The result in Theorem
\ref{thm:uniq} is actually more powerful in that it states
that there is a unique functor from $\underline{\mathcal
M}^{gen}$ to $\underline{\mathcal P}$ that satisfies
certain natural conditions.} }
\subsection{Complete and Average linkage are not functorial over $\mmon$}
It is interesting to point out
why complete linkage and average linkage (agglomerative) clustering,
as defined in Example \ref{hierarchical} fail to be functorial on
$\underline{\mathcal M}^{inj}$ (and therefore also on $\mgen$ as well) . A simple example explains this: consider
the metric spaces $X=\{A,B,C\}$ with metric given by the edge lengths
$\{4,3,5\}$ and $Y=(A',B',C')$ with metric given by the edge lengths
$\{4,3,2\}$, as given in Figure
\ref{cexample}. Obviously the map $f$ from $X$ to $Y$ with $f(A)=A'$,
$f(B)=B'$ and $f(C)=C'$ is a morphism in $\underline{\mathcal
M}^{inj}$. Note that for example for $r=3.5$ (shaded regions of the
dendrograms in Figure \ref{cexample}) we have that the partition of
$X$ is $\Pi_X = \{\{A,C\},B\}$ whereas the partition of $Y$ is $\Pi_Y
= \{\{A',B'\},C'\}$ and thus $f^*(\Pi_Y)=\{\{A,B\},\{C\}\}$. Therefore
$\Pi_X$ does not refine $f^*(\Pi_Y)$ as required by functoriality. The same
construction yields a counter-example for average linkage.
\begin{figure}[htb]
\centering
\includegraphics[width=.7\linewidth]{triang_rearr.eps}
\caption{An example that shows why complete linkage
fails to be functorial on $\underline{\mathcal M}^{inj}$ (and therefore also on $\mgen$).}
\label{cexample}
\end{figure}
\subsection{Functoriality over $\mgen$: a uniqueness theorem}
We prove a theorem of the same flavor as the
main theorem of \cite{kleinberg}, except that we prove existence and
uniqueness on $\underline{\mathcal M}^{gen}$ instead of impossibility
in our context.
\begin{theorem} \label{thm:uniq} Let $\hfunc: \underline{{\mathcal M}}^{gen} \rightarrow \underline{{\mathcal P}} $
be a hierarchical clustering functor which satisfies the following conditions.
\begin{description}
\item[(I)] {Let $\alpha : \underline{{\mathcal M}}^{gen} \rightarrow
\underline{Sets}$ and $\beta :\underline{ {\mathcal P}}\rightarrow
\underline{Sets}$ be the forgetful functors $(X,d_X) \rightarrow
X$ and $(X,\theta_X) \rightarrow X$, which forget the metric and
persistent set respectively, and only ``remember'' the underlying sets
$X$. Then we assume that $\beta \compcirc \hfunc = \alpha$. This
means that the underlying set of the persistent set associated to
a metric space is just the underlying set of the metric space. }
\item[(II)] {For $\delta\geq 0$ let
$\Delta_2(\delta)=(\{p,q\},\dmtwo{\delta})$ denote the two point
metric space with underlying set $\{ p,q \}$, and where
$\mbox{dist}(p,q) = \delta$. Then $\hfunc (\Delta_2(\delta))$ is the
persistent set $(\{p,q\},\theta_{\Delta_2(\delta)})$ whose
underlying set is $\{p,q \}$ and where $\theta_{\Delta_2(\delta)}
(t) $ is the partition with one element blocks when $t <
\delta$ and is the partition with a single two point block
when $t \geq \delta$. }
\item[(III)] { Write
$\hfunc(X,d_X)=(X,\theta^{\hfunc})$, then for any $t <
\mbox{sep}(X)$, the partition $\theta^{\hfunc}(t)$ is the
discrete partition with one element blocks. }
\end{description}
Then $\hfunc $ is equal to the functor $\rfunc$.
\end{theorem}
We should point out that another characterization of single linkage
has been obtained in the book \cite{jardine-sibson}.
\begin{proof}[Proof of Theorem \ref{thm:uniq}]
Write $\hfunc(X,d_X) = (X,\theta_X^{\hfunc})$. For each $r\geq 0$ we will
prove that \textbf{(a)} $\theta_X^{{ \mathrm{VR}}}(r)$ is a refinement of $\theta_X^{\hfunc}(r)$
and \textbf{(b)} $\theta_X^{\hfunc}(r)$ is a refinement of $\theta_X^{{ \mathrm{VR}}}(r)$. Then it will follow that $\theta_X^{{ \mathrm{VR}}}(r)=\theta_X^{\hfunc}(r)$ for all
$r\geq 0$, which shows that the objects are the same. Since this is a situation where, given any pair of objects, there is at most one morphism between them, this also determines the effect of the functor on morphisms.
Fix $r\geq 0$. In order to obtain (a) we need to prove that whenever
$x,x'\in X$ lie in the same block of the partition $\theta_X^{{ \mathrm{VR}}}(r)$,
that is $x\sim_r x'$, then they both lie in the same block of
$\theta_X^{\hfunc}(r)$.
skip
It is enough to prove the following
\begin{claim} Whenever
$d_X(x,x')\leq r$ then $x$ and $x'$ lie in the same block of $\theta_X^{\hfunc}(r)$.
\end{claim}
Indeed, if the claim is true, and $x\sim_r x'$ then one can find
$x_0,x_1,\ldots,x_n$ with $x_0=x$, $x_n=x'$ and $d_X(x_i,x_{i+1})\leq
r$ for $i=0,1,2,\ldots,n-1$. Then, invoking the claim for all pairs
$(x_i,x_{i+1})$, $i=0,\ldots,n-1$ one would find that: ${x}=x_0$ and
$x_1$ lie in the same block of $\theta_X^{\hfunc}(r)$, $x_1$ and
$x_2$ lie in the same block of $\theta_X^{\hfunc}(r)$, $\ldots$, $x_{n-1}$
and $x_n=x'$ lie in the same block of $\theta_X^{\hfunc}(r)$. Hence, $x$
and $x'$ lie in the same block of $\theta_X^{\hfunc}(r)$.
So, let's prove the claim. Assume $d_X(x,x')\leq r$, then the function
given by $p \rightarrow x, q \rightarrow x^{\prime}$ is a morphism
$g:\Delta_2(r) \rightarrow ( X, d_X) $ in $\underline{{\mathcal M}}^{gen}$.
This means that we obtain a morphism $$\hfunc (g) : \hfunc (\Delta_2(r))
\rightarrow \hfunc (X,d_X)$$ in $\underline{{\mathcal P}}$. But, by assumption (\textrm{II}), $p$ and
$q$ lie in the same block of the partition $\theta_{\Delta_2(r)}(r)$. By
functoriality it follows that
$\hfunc(g)$ is persistence preserving
and hence the elements $g(p) = x$ and $g(q) = x^{{\prime}}$ lie in the
same block of $ \theta_X^{\mathfrak{H}} (r)$. This concludes the proof of (a).
skip
For condition (b), assume that $x$ and $x'$ belong to the same block
of the partition $\theta_X^{\hfunc}(r)$ for some $r\geq 0$. We will prove that necessarily
$x\sim_r x'$. This of course will imply that $x$ and $x'$ belong to the
same block of $\theta_X^{{ \mathrm{VR}}}(r)$.
skip
Consider the metric space $(X_r,d_{r})$ whose points are the
equivalence classes of $X$ under the equivalence relation $\sim _r$,
and where the distance $d_{r}({\mathcal B},{\mathcal B'})$ between
two equivalence classes ${\mathcal B}$ and ${\mathcal B}' $ is defined to be
the maximal ultra-metric pointwisely less than or equal to
$$W_X(\mathcal{B},\mathcal{B}'):= \min_{x
\in {\mathcal B}}\min _{x^{\prime} \in {\mathcal B}'} d_X
(x,x^{\prime}),$$
that is $d_r=\mathcal{U}(W_X)$, see (\ref{eq:U}). It follows from the definition of $\sim _r$ and Lemma \ref{prop:props-uaster}
that if two equivalence classes are distinct, then the distance
between them is $> r$. Indeed, for otherwise there would be $x,x'\in X$ with $[x]_r\neq [x']_r$ and $d_r([x]_r,[x']_r)\leq r$, which would imply that $W_X([x]_r,[x']_r)\leq r$, which is a contradiction. Hence $\mbox{sep}(X_r,d_r)>r.$
Write $\hfunc (X_r,d_{r})=(X_r,\theta_{X_r}^\hfunc)$. Since
$\mbox{sep}(X_r)>r$, hypothesis (III) now directly shows that the
blocks of the partition $\theta_{X_r}^\hfunc(r)$ are exactly the equivalence
classes of $X$ under the equivalence relation $\sim _r$, that is
$\theta_{X_r}^\hfunc(r) = \theta_X^{{ \mathrm{VR}}}(r)$. Finally, consider the morphism
$$\pi_r: (X,d_X)
\rightarrow (X_r,d_{r})$$ in $\underline{{\mathcal M}}^{gen}$ given on elements $x
\in X$ by $\pi_r (x) =[x]_r$, where $[x]_r$ denotes the equivalence class of
$x$ under $\sim_r$. By functoriality, $\hfunc(\pi_r):(X,\theta_X^{\hfunc})\rightarrow (X_r,\theta_{X_r}^\hfunc)$ is
persistence preserving, and therefore, $\theta^{\hfunc}(r)$ is a
refinement of $\theta_{X_r}(r)=\theta_X^{{ \mathrm{VR}}}(r)$. This is depicted as follows:
\begin{displaymath}
\xymatrix{
(X,d_X) \ar[r]^\hfunc \ar[d]_{\pi_r} & (X,\theta_X^{{ \mathrm{VR}}}) \ar[d]^{\hfunc(\pi_r)} \\
(X_r,d_{r}) \ar[r]^{\hfunc} & (X_r,\theta_{X_r}^\mathfrak{H}) }
\end{displaymath}
This concludes the proof of (b).
\end{proof}
\subsubsection{Comments on Kleinberg's conditions}\label{sec:klein-conds}
\noindent We conclude this section by observing that analogues of the three (axiomatic)
properties considered by Kleinberg in \cite{kleinberg} hold for $\rfunc$.
\begin{itemize}
\item Kleinberg's {first condition} was {\em scale-invariance}, which
asserted that if the distances in the underlying point cloud
data were multiplied by a constant positive multiple
$\lambda$, then the resulting clustering decomposition should
be identical. In our case, this is replaced by the condition (Recall notation of Example \ref{scaling})
that ${\mathfrak R}\compcirc\sigma _{\lambda}(X,d_X) =
s_{\lambda}\compcirc{\mathfrak R}(X,d_X)$, which is
trivially satisfied.
\item Kleinberg's {second condition}, {\em richness}, asserts that
any partition of a dataset can be obtained as the result of
the given clustering scheme for some metric on the
dataset. In our context, partitions are replaced by
persistent sets. Assume that there exist $t\in \R_+$
s.t. $\theta_X(t)$ is the single block partition, i.e., impose
that the persistent set is a dendrogram (cf. Definition
\ref{def:persistent-sets}). In this case, it is easy to
check that any such persistent set can be obtained as ${\mathcal
R}^{gen}$ evaluated for some (pseudo)metric on some
dataset. Indeed,\footnote{We only prove triangle inequality.}
let $(X,\theta_X)\in{ob}(\underline{\mathcal P})$. Let
$\epsilon_1,\ldots,\epsilon_k$ be the (finitely many)
transition/discontinuity points of $\theta_X$. For $x,x'\in X$
define $d_X(x,x')=\min\{\epsilon_i\}$ s.t. $x,x'$ belong to
same block of $\theta_X(\epsilon_i)$.
This is a pseudo metric on $X$. Indeed, pick points $x,x'$
and $x''$ in $X$. Let $\epsilon_1$ and $\epsilon_2$ be
minimal s.t. $x,x'$ belong to the same block of
$\theta_X(\epsilon_1)$ and $x',x''$ belong to the same block of
$\theta_X(\epsilon_2)$. Let
$\epsilon_{12}:=\max(\epsilon_1,\epsilon_2)$. Since
$(X,\theta_X)$ is a persistent set (Definition
\ref{def:persistent-sets}), $\theta_X(\epsilon_{12})$ must have
a block $\mathcal B$ s.t. $x,x'$ and $x''$ all lie in
$\mathcal B$. Hence $d_X(x,x'')\leq \epsilon_{12}\leq
\epsilon_1+\epsilon_2=d_X(x,x')+d_X(x',x'')$.
\item Finally, Kleinberg's {third condition},
\emph{consistency}, could be viewed as a rudimentary example
of functoriality. His morphisms are similar to the ones in
$\underline{\mathcal M}^{gen}$.
\end{itemize}
\subsection{Functoriality over $\mmon$}
In this section, we illustrate how relaxing the functoriality permits
more clustering algorithms. In other words, we will restrict ourselves
to $\mmon$ which is smaller (less stringent) than
$\mgen$ but larger (more stringent) than
$\miso$.
Fir a finite metric space $(X,d_X)$ and $r\geq 0$. For $x \in X$, let $[x]_r$ be the equivalence class of $x$ under the
equivalence relation $\sim _r$, and define $c(x) = \big|[x]_r\big|$. For any positive
integer $m$, we now define $X_m \subseteq X$ by $X_m = \{ x \in X;\
c(x) \geq m\} $. We note that for any morphism $f: X \rightarrow Y$
in $\mmon$, we find that $f(X_m) \subseteq Y_m$.
This property clearly does not hold for more general morphisms. For
every $r$, we can now define a new equivalence relation $\sim ^m_r $
on $X$, which refines $\sim _r$, by requiring that each equivalence
class of $\sim _r$ which has cardinality $\geq m$ is an equivalence
class of $\sim _r^m$, and that for any $x $ for which $c(x) < m$, $x$
defines a singleton equivalence class in $\sim ^m_r$. We now obtain a
new persistent set $(X,\theta_X^m )$, where $\theta_X ^m (r)$ will denote
the partition associated to the equivalence relation $\sim ^m_r$. It
is readily checked that $X \rightarrow (X, \theta_X^m )$ is functorial
on $\mmon$.
\begin{definition}
For each $m\in\N$ let $\hfunc^{m}:\mmon\rightarrow \catp$ be the functor which to each finite metric space $(X,d_X)$ assigns the persistent set $(X,\theta^m_X)$ where for each $r\geq 0$, $\theta_X^m(r)$ is the partition of $X$ into equivalence classes of $\sim_r^m$.
\end{definition}
This scheme could be motivated by
the intuition that one does not regard clusters of small cardinality
as significant, and therefore makes points lying in small clusters
into singletons, where one can then remove them as representing
``outliers'' or ``noise''.
Another construction arises from similar consideration to those that gave rise to the standard clustering algorithms $\cfunc^{\Delta_m(\delta)}$ in Section \ref{section:mmon-density}, cf. Corollary \ref{coro:change-metric}.
\begin{definition}
For each $m\in\N$ define the functor $\rfunc^{\Delta_m}:\mmon\rightarrow\catp$ given by
$$\rfunc^{\Delta_m} = \rfunc\compcirc \tfunc^m,$$
where we have written $\tfunc^m = \tfunc^\Omega$ for $\Omega= \{\Delta_m(\delta),\ \delta\geq 0\}$, cf. (\ref{eq:Tfunc}).
\end{definition}
That this definition produces a functor follows directly from the fact that $\rfunc$ and $\tfunc^m$ are functorial.
The resulting construction is in the same spirit as DBSCAN \cite{dbscan}.
\begin{figure}[htb]
\centering
\includegraphics[width=1\linewidth]{fig-CDeltam-no-part.eps}
\caption{Dendrograms resulting from applying $\rfunc^{\Delta_m}$ to a randomly generated dataset consisting of 50 points in the plane. The dataset consists of two clusters of points which are joined by a thin chain of points with some additional outliers. On the top row, from left to right we show the dendrograms corresponding to $m=2,3$ and $4$. Note how increasing $m$ produces dendrograms that exhibit more clearly clustered subsets.} \label{fig:cDeltam}
\end{figure}
\begin{remark} Notice first that $\rfunc^{\Delta_2}=\rfunc$. Also,
for each $m\in\N$, the hierarchical clustering functor $\rfunc^{\Delta_m}$ is related to the standard clustering functor $\cfunc^{\Delta_m(\delta)}$ in the following manner. Fix $\delta\geq 0$ and write $\cfunc^{\Delta_m(\delta)}(X,d_X)=(X,P_X)$, then, $P_X=\theta_X^m(\delta).$
\end{remark}
An example of dendrograms obtained with these methods is shown in Figure \ref{fig:cDeltam}.
\section{Discussion} \label{various}
The methods of this paper permit natural extensions to other situations,
such as clustering of graphs and networks with, for example, the notion of clique clustering
\cite{hungarian-clique} fitting naturally into our context. It appears likely that from the point of view described
here, it will in many cases be possible, given a collection of
constraints on a clustering functor, to determine the {\em universal }
one satisfying the constraints. One could therefore use sets of
constraints as the definition of clustering functors.
In addition to
producing classifications of clustering functors, functoriality is a very
desirable property when clustering is used as an ingredient in
computational topology. Functorial clustering schemes produce diagrams of
sets of varying shapes which can be used to construct simplicial complexes
which in turn reflect the topology of connected components \cite{carlsson-09}. Functorial schemes also can be used to construct ``zig-zag diagrams"
\cite{zigzag} which reflect the stability of
clusters produced over a family of independent samples.
We believe that the conceptual framework presented here can be a
useful tool in reasoning about clustering algorithms. We have also
shown that clustering methods which have some degree of functoriality
admit the possibility of certain kind of qualitative geometric
analysis of datasets which can be quite valuable. The general idea
that the morphisms between mathematical objects (together with the
notion of functoriality) are critical in many situations is
well-established in many areas of mathematics, and we would argue that
it is valuable in this statistical situation as well.
\begin{thebibliography}{EpKSX96}
\bibitem[Car09]{carlsson-09}
Gunnar Carlsson.
\newblock Topology and data.
\newblock {\em Bull. Amer. Math. Soc.}, 46:255--308, 2009.
\bibitem[CdS10]{zigzag}
G.~Carlsson and V.~de~Silva.
\newblock Zigzag persistence.
\newblock {\em Foundations of computational mathematics}, 10(4), August 2010.
\bibitem[CM08]{carlsson-memoli-full}
G.~{Carlsson} and F.~{M\'emoli}.
\newblock {Persistent Clustering and a Theorem of J. Kleinberg}.
\newblock {\em ArXiv e-prints}, August 2008.
\bibitem[CM10a]{clust-um}
Gunnar Carlsson and Facundo M\'emoli.
\newblock Characterization, stability and convergence of hierarchical
clustering methods.
\newblock {\em Journal of Machine Learning Research}, 11:1425--1470, 2010.
\bibitem[CM10b]{multi-param}
Gunnar Carlsson and Facundo M\'emoli.
\newblock Multiparameter clustering methods.
\newblock In Claus~Weihs Hermann Locarek-Junge, editor, {\em Classification as
a Tool for Research. Proc. 11th Conference of the International Federation of
Classification Societies (IFCS-09)}, pages 63--70, Heidelberg-Berlin, 2010.
Springer-Verlag.
\bibitem[EpKSX96]{dbscan}
Martin Ester, Hans peter Kriegel, J\"org S, and Xiaowei Xu.
\newblock A density-based algorithm for discovering clusters in large spatial
databases with noise.
\newblock pages 226--231. AAAI Press, 1996.
\bibitem[GvLW09]{science_art}
I.~Guyon, U.~von Luxburg, and R.~Williamson.
\newblock Clustering: Science or art?
\newblock Technical report, Paper presented at the NIPS 2009 Workshop
Clustering: Science or Art?\\
\url{http://www.kyb.tuebingen.mpg.de/bs/people/ule/publications/publication_downloads/IsClusteringScience.pdf}, 2009.
\bibitem[Hau95]{hausmann}
Jean-Claude Hausmann.
\newblock On the {V}ietoris-{R}ips complexes and a cohomology theory for metric
spaces.
\newblock In {\em Prospects in topology ({P}rinceton, {NJ}, 1994)}, volume 138
of {\em Ann. of Math. Stud.}, pages 175--188. Princeton Univ. Press,
Princeton, NJ, 1995.
\bibitem[Isb64]{isbell}
J.~R. Isbell.
\newblock Six theorems about injective metric spaces.
\newblock {\em Comment. Math. Helv.}, 39:65--76, 1964.
\bibitem[JD88]{clusteringref}
Anil~K. Jain and Richard~C. Dubes.
\newblock {\em Algorithms for clustering data}.
\newblock Prentice Hall Advanced Reference Series. Prentice Hall Inc.,
Englewood Cliffs, NJ, 1988.
\bibitem[JS71]{jardine-sibson}
Nicholas Jardine and Robin Sibson.
\newblock {\em Mathematical taxonomy}.
\newblock John Wiley \& Sons Ltd., London, 1971.
\newblock Wiley Series in Probability and Mathematical Statistics.
\bibitem[Kle02]{kleinberg}
Jon~M. Kleinberg.
\newblock An impossibility theorem for clustering.
\newblock In Suzanna Becker, Sebastian Thrun, and Klaus Obermayer, editors,
{\em NIPS}, pages 446--453. MIT Press, 2002.
\bibitem[LW67]{lance67general}
G.~N. Lance and W.~T. Williams.
\newblock A general theory of classificatory sorting strategies 1.
{Hierarchical} systems.
\newblock {\em Computer Journal}, 9(4):373--380, February 1967.
\bibitem[ML98]{maclane}
Saunders Mac~Lane.
\newblock {\em Categories for the working mathematician}, volume~5 of {\em
Graduate Texts in Mathematics}.
\newblock Springer-Verlag, New York, second edition, 1998.
\bibitem[Mun75]{munkres}
James~R. Munkres.
\newblock {\em Topology: a first course}.
\newblock Prentice-Hall Inc., Englewood Cliffs, N.J., 1975.
\bibitem[NS08]{handbook-um}
Amiya Nayak and Ivan Stojmenovic.
\newblock {\em Handbook of Applied Algorithms: Solving Scientific, Engineering,
and Practical Problems}.
\newblock Wiley-IEEE Press, 2008.
\bibitem[PDFV05]{hungarian-clique}
Gergely Palla, Imre Derenyi, Illes Farkas, and Tamas Vicsek.
\newblock Uncovering the overlapping community structure of complex networks in
nature and society.
\newblock {\em Nature}, 435(7043):814--818, June 2005.
\end{thebibliography}
\end{document}
|