This is a web version of a previously published paper. The text is reproduced here in its entirety but no figures are included. Editorial insertions in this web version are enclosed in {braces}. Citations to this paper should refer to the original printed version:

Robinson, Peter M.W., & Robert J. O’Hara. 1996.
Cladistic analysis of an Old Norse manuscript tradition. Research in Humanities Computing, 4: 115–137.

Cladistic Analysis of an Old Norse Manuscript Tradition

Peter M.W. Robinson
Computers and Manuscripts Project
Oxford University Computing Services
13 Banbury Road, Oxford OX2 6NN U.K.
Robert J. O’Hara
Department of Philosophy and The Zoological Museum
University of Wisconsin–Madison
Madison, Wisconsin 53706 U.S.A.

1. Introduction

The Old Norse narrative sequence Svipdagsmál, consisting of the poems Gróugaldr and Fjölsvinnsmál, together about 1500 words in length, is extant in forty-six manuscripts written between 1650 and 1830. In preparing his edition of Svipdagsmál, Robinson (1991) transcribed and collated forty-three of these manuscripts by computer, and stored the results of the collation in a relational database.1 From close examination of the distribution of the variant readings as stored in the database, and from the relatively ample external evidence concerning the history of these late manuscripts, a stemma, or tree of the historical relationships of the manuscripts, was constructed (Fig. 1). The construction of manuscript stemmata such as this is one of the central challenges in the field of textual criticism.

The analysis of the collation of Svipdagsmál using the database and the supplementary information obtained from examination of the manuscripts themselves was a clumsy and laborious process that occupied some six months, during which hypotheses concerning the relationships among the manuscripts were framed, tested in the database, modified, retested, and so on (Robinson 1989a, 1989b). A similar method of textual analysis, also using database techniques, has been described by Xhardez (1994). The task of collation itself, for Svipdagsmál and similar large manuscript traditions, has now been greatly eased by the redevelopment of Robinson’s collation programs in the form of the package Collate, now generally available for Macintosh computers (Robinson 1994). The problem of reconstructing the stemmatic relationships among the manuscripts of a large tradition, however, has remained exceedingly difficult.

The difficulty of reconstructing manuscript stemmata, and the highly structured character of the data that result from collation, have suggested to a number of authors that statistical clustering techniques might prove valuable in pointing quickly to possible relationships which could then be thoroughly examined by other means. The first scholar to have tried such a statistical approach appears to have been John Griffith, who applied cluster analysis to some variant readings in the manuscripts of Juvenal (Griffith 1968; Griffith 1984). Most subsequent studies of this type have also used either cluster analysis or multivariate analysis (see the reviews in Hockey 1980 and Pierce 1988). Only one of these studies—that of Xhardez (1994) on some fifty manuscripts of a twelfth century text—attempted to apply statistical techniques to all the data from a complete manuscript tradition. Xhardez found that this gave a ‘general but fairly accurate idea of the broad relationships between the manuscripts’.

In this paper we report on the preliminary application of a different technique to the reconstruction of manuscript histories, using the Svipdagsmál tradition as an example. This technique is known as cladistic analysis, and it has been developed over the last thirty years by researchers in the field of systematics, the branch of evolutionary biology that specializes in the reconstruction of the evolutionary tree of life. Although it makes use of computer-assisted analysis of numerically coded data, just as the above-mentioned clustering techniques do, the object of cladistic analysis is quite different from the object of clustering. Rather than attempting to assemble groups of similar objects, cladistic analysis attempts to reconstruct the history of objects that are related in a tree of ancestry and descent. Because of their explicitly historical character (O’Hara, 1988a), the concepts and methods of cladistic analysis can be readily adapted to the reconstruction of manuscript stemmata.

2. Trees of History in Systematics and in Textual Criticism

There are at least three different disciplines that reconstruct what might be called ‘trees of history’: systematic biology, historical linguistics, and stemmatics. The similarities between two of these disciplines—systematic biology and historical linguistics—have been noted at least since the mid-nineteenth century, and Charles Darwin conjectured in the Origin of Species (1859, 422) that the tree of human languages would correspond to the evolutionary tree of the human races. Rulon Wells (1987), in a pioneering interdisciplinary volume edited by Hoenigswald and Wiener (1987), has provided an excellent survey of the comparisons between linguistics and systematics that were made by early workers in both fields.

The similarities between systematic biology and stemmatics, in contrast, seem to have been noticed only recently (Platnick and Cameron 1977; Cameron 1987; Lee 1989), in spite of the fact that the two fields have had remarkably parallel histories for the last two hundred years. As he began to develop the rudiments of his theory of evolution in the 1830s, Charles Darwin began to sketch trees of evolutionary descent in his research notebooks (Barrett et al. 1987, 176–182), and these culminated in the theoretical diagram of descent that appeared in 1859 as the only illustration in the Origin of Species (Fig. 2). At very nearly the same time Karl Zumpt published the first tree of manuscript descent, a stemma (Fig. 3) that appeared in the introduction to his edition of Cicero’s Verrine Orations (Zumpt 1831).3 Zumpt’s work was followed a number of years later by Karl Lachmann’s edition of Lucretius (Lachmann 1850), a highly influential study that has associated Lachmann’s name with stemmatics, which is now often referred to as ‘the Lachmann method’.

The formulations of Darwin, Zumpt, and Lachmann were of course preceded by long intellectual prehistories. In systematics, the idea of ‘the natural system’—the arrangement or classification of the totality of life—had been important since the seventeenth century (Sloan 1972; Winsor 1976; Stevens 1982, 1984; O’Hara 1988b, 1991), and the natural system was often represented diagrammatically in the early systematic literature as a map, or a network, or a set of circles, or a tree. It is important to realize, however, that in these early works the notion of the natural system was ordinarily not an historical or evolutionary notion, and that the systematic trees of the pre-evolutionary period were not trees of history or genealogy, but rather ahistorical arrangements or logical classificatory devices. The conceptual transformation in systematics that began with Darwin’s diagrams in the early 1800s was the modification of the tree of logical relationships into a genealogical tree of historical relationships, a modification that did not take place in a completely orderly manner, and which in fact has only been completed quite recently (see further below). In the history of manuscript studies, Politian determined in 1489 the relationships of a number of manuscripts of Cicero’s Letters by recognizing a significant error in one copy that also appeared in several other, later copies, suggesting that all these later copies had descended from that one ancestor (Timpanaro 1963, 4–5). The great English scholar Richard Bentley (1662–1742) established the aim of historical editing, declaring in a letter to the Archbishop of Canterbury in 1716 that he could restore the text of the Greek New Testament so that it would be ‘exactly as it was in the best exemplars at the time of the council of Nice’ (Wordsworth 1842, 503). In the century following Bentley’s grand assertion other scholars laid the foundations of historical reconstruction of the history of manuscript traditions (Bengel 1763, 18–20; Semler 1765, 88–99; Griesbach 1796). Finally, Lachmann (1876, II 276) proclaimed the triumph of stemmatics: ‘the establishment of a text according to its tradition is a strictly historical undertaking’.4 Sebastian Timpanaro’s studies on the history of the Lachmann method (1963, 1971, 1981) trace in detail the development of this historical or ‘genetic’ approach to the study of manuscripts.

Once the tree model became established in both manuscript studies and in systematics, it had a great impact. Textual editors thought they had found the key to unlock the secrets of large, difficult and important manuscript traditions; natural historians thought they had found the key to life itself. In the years between 1850 and 1930 textual scholars applied stemmatic techniques widely, and drew trees showing the histories of many manuscript traditions. In a famous essay on the manuscript tradition of the Old French ‘Lai de l’Ombre’, Joseph Bédier (1928) stated that he had examined over a hundred manuscript trees that had been drawn up by scholars. Similarly, evolutionary biologists from 1860 to 1900 constructed many trees showing the evolutionary history of animals and plants (Voss 1952; Reif 1983; Stevens 1984; Oppenheimer 1987; O’Hara 1988b, 1991; Craw 1992).

The burst of enthusiasm for tree making in both manuscript studies and in systematics in the late 1800s was followed by a period of disillusionment, however. Around the turn of the twentieth century, some biologists began to criticize the practice of making trees, calling it a ‘speculative’ activity, and many workers began to turn toward either populational studies or experimental studies of proximate causation, and away from large-scale historical investigations (O’Hara 1988b, 1991; but see Craw 1992). In textual scholarship at roughly the same time a series of assaults on stemmatics began to shake it to its roots. Bédier’s essay, referred to above, was particularly influential. He pointed out that 105 of the 110 manuscript trees he had surveyed were ‘bifid’ trees, with all the manuscripts in every tradition descending in just two main branches from the hypothesized ancestor. This meant that editors concerned with reconstructing the ancestral text could regard each branch of the tree as of equal potential authority, and when faced with conflicting readings could choose whichever they liked as if the tree did not exist. The suspicion was that the trees had been constructed just so they could be ignored, and final readings chosen to suit the individual taste of the editors (see also Timpanaro 1963, 112–35; Alberti 1979; but cf. Reeve 1986). Other scholars discovered that the practical creation of a single tree from a confusion of manuscripts was, simply, impossible. Kane and Donaldson found this with the manuscripts of the A and B versions of Langland’s Piers Plowman (Kane 1960; Kane and Donaldson 1975); Dawe (1964) with the manuscripts of Aeschylus; Bévenot with the manuscripts of St. Cyprian (1961). Maas (1949, 1958) attempted to establish a formal set of rules for the construction of stemmata, but found he could do so only by setting aside the question of scribal borrowing from manuscripts other than the exemplar (borrowing known as ‘contamination’ or ‘horizontal transmission’). Pasquali criticized Maas’s rules in a much longer book (1952) showing just how widespread contamination is. When faced with real variants, in real manuscripts, with substantial contamination from manuscript to manuscript, with scribes emending away errors or independently introducing the same error in different manuscripts, stemmatics as traditionally practiced was seen to collapse into absurdity. In 1939, Vinaver pronounced the death of stemmatics: ‘the ingenious technique of editing evolved by the great masters of the nineteenth century has become as obsolete as Newton’s physics’ (1939, 351). Textual editing did continue of course: editors produced either ‘single-text’ editions, as Bédier proposed, such as the many editions of the Canterbury Tales based either on the Hengwrt or Ellesmere manuscripts, or ‘eclectic’ texts, with readings drawn piecemeal from all manuscripts as the editors’ taste, knowledge, and intuition dictated, such as Kane and Donaldson’s work on Langland. But in these cases the evolutionary history of the text, as it was copied from manuscript to manuscript, was essentially irrelevant, and these editions existed in an historical void.

3. Systematics Since 1965

While the situation in textual scholarship remains today largely as it has been for much of this century, the situation in systematic biology has changed radically since 1965. Out of the interaction of several different trends beginning in the mid-1960s has come the approach to systematics known as cladistics, and it has swept the field since the 1970s. We describe the recent history of systematics in some detail here in order to demonstrate clearly how different systematics is today from when textual scholars, such as Griffith (1968), last looked in on it.

The first of the interacting elements that led to the development of contemporary cladistics was the numerical approach to systematics developed in the 1950s and 1960s by the now-defunct ‘phenetic’ school (Sokal and Sneath 1963; Sneath and Sokal 1973). Some members of the phenetic school argued that evolutionary history was virtually unknowable, and that systematists ought to abandon the search for phylogenetic trees, and instead simply produce practical and efficient classifications of organisms based on as many organismal characteristics as possible. These classifications were viewed as operational devices only, with no particular relation to the evolutionary history of the organisms they classified. The phenetic approach was in a sense the last gasp of those who viewed historical studies as ‘speculative’, and although it appealed to some systematists for a time, it was widely criticized on both practical and theoretical grounds (Mayr 1965; Hull 1970; Johnson 1970; Farris 1977). As interest in phylogenetic reconstruction blossomed again in the 1960s and 1970s, the phenetic approach to systematics came to meet with near-universal rejection. The phenetic school did have some lasting salutary effects, however: it broke down much of the resistance to quantitative analysis that had existed in the field previously, and it introduced the use of computers in systematic data analysis.

The second trend that contributed to the development and adoption of cladistic methods was the translation and study of a work entitled Phylogenetic Systematics, by the entomologist Willi Hennig. Originally published in German in 1950 (see Craw 1992 for a thorough study of the impact of the German edition), this work was revised and published in English, first in abstract in 1965, and then fully in 1966. Many of the explicitly articulated principles of phylogenetic reconstruction that Hennig’s volume contained came to be adopted by the systematics community, although not without a good deal of tumultuous debate (Hull 1988). A number of workers who had originally been associated with the phenetic school began to apply their quantitative skills to the problems of phylogeny reconstruction, and eventually developed quantitative formulations of Hennig’s principles, and the similar principles of W.H. Wagner (1961, 1980). These formulations came to be referred to as numerical cladistics or quantitative phyletics (Camin and Sokal 1965; Kluge and Farris 1969; Farris, Kluge, and Eckhardt 1970), the terms ‘cladistics’, ‘phyletics’, and ‘phylogenetics’ being used to clearly distinguish work that aimed at reconstructing branching evolutionary history (phylogeny) from phenetic work which did not.

The third contributory element in the development of contemporary cladistic systematics was the availability of new information on molecular variation among organisms. Comparative data on the sequential structure of protein molecules and on gene frequencies within populations became available during the 1960s, and a number of molecular biologists began to investigate the possibility of using these data to reconstruct evolutionary trees (Edwards and Cavalli-Sforza 1964; Zuckerkandl and Pauling 1965; Fitch and Margoliash 1967). Characteristic of the work of these investigators and their followers were attempts to create models of the underlying process of descent and change, and to use these models in choosing among differing estimates of the true phylogeny. Work in this area has become increasingly sophisticated as the structure and function of the molecular machinery of heredity has become better understood.

Through the 1960s and 1970s these different movements interacted and reacted to one another, and something of a synthesis took place, with some ideas from each being retained and others discarded. This is of course only a capsule summary of a very complicated history, but it gives some sense of how dynamic systematics has been in recent years, and how different it is today from the way it was only a few years ago. As in any academic field, many details of cladistic theory and practice remain contentious, and lively theoretical debates continue in journals such as Systematic Zoology (now Systematic Biology), Cladistics, and Systematic Botany. In spite of some continuing controversy the field has now stabilized considerably, hundreds of new evolutionary trees or ‘cladograms’ are published every year for a great variety of organisms, and a number of different computer programs are available that implement cladistic procedures.

These three developments in systematics—advent of numerical data analysis, the theoretical foundations laid by Hennig and others, and the mass of information provided by molecular sequences—are all relevant to our belief that cladistic methods are applicable to stemmatics. Firstly, the computer programs now used by cladists may also be used by stemmaticists. Secondly, the theoretical rationale for reconstructing the evolutionary history of a collection of organisms by analysis of their characteristics parallels stemmaticists’ belief that the history of a manuscript tradition may be recovered from analysis of variant readings. Thirdly, computer collation can provide great quantities of information about how manuscripts agree and disagree, similar to the information supplied to evolutionary biologists by molecular sequencing.

4. The Principles of Cladistic Analysis

What are the principles of cladistic analysis as they are used by systematic biologists? The diversity of life is of staggering magnitude: several million species of organisms have been described, and probably several times that number still await description. When we examine this diversity, we see a great variety of differences among organisms: differences in size, colour, growth, behaviour, external anatomy, internal anatomy, physiological responses, ecological preferences, molecular makeup, and so on. These differences have arisen and accumulated through the long course of reproduction and divergence that makes up the evolutionary past. They have arisen, in Darwin’s concise phrase, through ‘descent with modification’. Suppose we wish to reconstruct this history of descent and modification in some branch of the evolutionary tree, the species of woodpeckers, say, or bats, or bird’s nest fungi. First we must search for what systematists call characters, that is differences among the species under study that can divide them into two or more groups, and from which evolutionary events can be inferred. In the case of woodpeckers, for example, we will find that some of the 210 known species have four toes, while others have only three toes. It might seem that a character such as this, with two states, would give us evidence that there are two branches in the sought-for phylogeny: a four-toed branch and a three-toed branch. Reflection will show, however, that one of these states is likely to be the ancestral or primitive state, present in the ancestor of the whole group originally, and potentially retained unmodified anywhere. If, for example, the ancestor of all woodpeckers was four-toed (as we believe to be the case), and the three-toed state (called the derived state) arose and was passed on in one branch of the woodpecker tree, the collection of four-toed species would not themselves constitute a whole branch of the tree. The branch defined by the derived, three-toed state, would be nested within the whole tree, which elsewhere would exhibit the ancestral four-toed condition. Distinguishing the ancestral from the derived states of characters is called ‘polarity determination’ in the terminology of cladistics, and it is an essential step, because it is only the derived states of characters that identify branches of the evolutionary tree.

Once a collection of characters has been described for a group under study, and the polarity of those characters has been determined, the characters are, in a sense, ‘added up’ to yield an estimate of the phylogeny as a whole, one that accounts for the observed distribution of character states among the descendants in the simplest manner (Fig. 4). When the number of characters and the number of taxa (organisms under study) is large, and when some of the characters conflict with one another owing to convergence, this can be a difficult task, as the number of possible trees that must be evaluated for their fit to the data becomes enormous.5 It is here that computer programs are of assistance, and several tree evaluation programs have been written and are in wide use in the systematics community (see Mayr and Ashlock 1991, 320–21 for a listing). These programs typically search through the range of possible trees, and determine the minimum number of character state changes that could have occurred on each tree, given the particular data supplied. The tree or trees on which the fewest changes overall are required (the ‘shortest’ or ‘most parsimonious’ tree or trees) are taken to be the best estimates of the true history of taxa under study.

Of course there are many details of method that have not been mentioned here, and many difficulties. What does one do when the polarity of a character cannot be reliably determined, for example, or when one believes that a character can change readily in one direction but not in another direction? These problems make the application of the basic principles of cladistics more difficult, and they have been extensively discussed in the cladistic literature (for thorough introductory treatments of cladistic analysis see Wiley 1981; Felsenstein 1982; Sober 1988; Swofford and Olsen 1990; Brooks and McLennan 1991; and Mayr and Ashlock 1991, Chapter 11). Troublesome as they can be, however, these problems have not prevented the successful application of cladistic techniques to a wide variety of problems in evolutionary biology.

It should be apparent from this outline that there is a fundamental identity between cladistic systematics and stemmatics. Each discipline seeks to explain the existence of a varied collection of objects (manuscripts for stemmaticists, organisms for systematists) that has resulted from a sequence of branching descents over time from a common ancestor. Accordingly, the object of cladistic analysis is nearly identical with the object of stemmatics: the reconstruction of a tree of descent based on comparative observations of the descendants themselves. The cladistic principle that only the derived states of characters identify tree branches has long been a principle of stemmatics, first spelt out by Lejay in 1888 (reported in Kenney 1974, 135), and has been repeated by many others (Kane 1984, 209; West 1973, 32–3; cf. Quentin 1926, 61–96). The ability of cladistics to deal with ‘one-way variation’ (irreversible characters) and ‘sequential variation’ (multistate ordered characters) is similarly important, since textual scholars have long been familiar with errors (such as omissions) which prohibit restoration of the original, as well as with errors that must occur in a particular sequence (Maas 1958, 4). With these parallels in mind, we thought it possible that cladistics might do for textual criticism what it has done for systematics: it might restore history.

5. Cladistic Analysis of the Svipdagsmál Material

As mentioned above, the parallels between stemmatics and cladistic systematics have been noted recently by several authors. Platnick and Cameron outlined some of these parallels in the journal Systematic Zoology in 1977, and Cameron detailed them further in an excellent review published in 1987. Arthur Lee, in a paper presented to the 1987 Patristics conference in Oxford, seems to have been the first actually to apply cladistic techniques to a particular problem in manuscript studies: the relationships of some twenty-five manuscripts of Augustine’s Quaestiones in Heptateuchum (Lee 1989). Our own analysis6 of the Svipdagsmál sequence brings together three elements not previously joined in any study: firstly, all the data from a complete collation of an entire manuscript tradition (Robinson 1994); secondly, a powerful and flexible cladistics program (PAUP, ‘Phylogenetic Analysis Using Parsimony’; Swofford 1991); and thirdly, a wealth of external evidence about how the manuscripts are related, evidence which could be used to test the results of the cladistic analysis (Robinson 1991). This last element, the external evidence, was particularly crucial. Previous attempts at numerical analysis have rarely been able to do more than replicate earlier non-numerically derived conclusions (for example, Moorman 1982 duplicating the results of Manly and Rickert 1940). If the earlier conclusions were themselves unsound (and there is considerable doubt concerning Manly and Rickert’s work; see for example Kane 1984), then the later numerical efforts may not carry conviction. The sixteen manuscripts in Robinson’s Svipdagsmál stemma (Fig. 1) that are linked by arrows are those for which there is clear external evidence (typically scribal statements in the manuscript) that they are related as given. This external evidence provided the opportunity to judge decisively the validity or invalidity of the cladistic approach.

The collection of data analyzed by the cladistics program PAUP was in the form of a table representing all the agreements and disagreements in the Svipdagsmál manuscripts. Each row of the data table represented a manuscript; each column represented a reading. A ‘0’ in a column meant simply that the manuscript did not have that reading; a ‘1’ meant that it did. Thus:

  manuscript A  0 1 1
  manuscript B  0 0 0
  manuscript C  1 1 0

would indicate that reading number one was present in manuscript C, but absent in manuscripts A and B, reading number two was present in manuscripts A and C, but absent in manuscript B, and so on. No weighting of any kind was applied to the data, and no readings or groups of readings were excluded from the analysis even though there was clear evidence of substantial contamination and coincident variation.

Fig. 5 gives the family tree for the manuscripts created by PAUP in its run over the raw collation data.7 Comparison of this with Fig. 1 shows the accuracy with which PAUP replicated the outlines of Robinson’s stemma. Firstly, the sixteen manuscripts which external evidence showed as directly related to one another: each of these manuscripts is placed very close to its known relative, usually adjacent. Note for example the sequence St copied to 34 to 1870, or the three manuscripts 1689, 5, and 329, all written by one scribe: these are placed directly adjacent in Fig. 5 just as they are in Robinson’s stemma. Secondly, there are major groupings of manuscripts having relationships with one another and with key individual manuscripts. The cladistic analysis identified all these correctly. For example the Egilsson group manuscripts are placed very close to Ra in Fig. 5, with one of them, 1868, separated by just one node. Without PAUP, establishing the closeness of these manuscripts to Ra took considerable effort. Another example can be seen in the B manuscripts, the group on the right in Fig. 1. After much effort without PAUP, Robinson had decided that the B manuscripts all descended from O, with another manuscript, 1872, descending on a different branch from O. That is very much how PAUP’s analysis places them in Fig. 5, with O and the B manuscripts appearing as coordinate branches (called ‘sister clades’ in systematics), and with 1872 sister to O and the B manuscripts taken together. Consider too the three manuscripts 818b, 3633, and 6: Fig. 1 shows these forming a sub-group of their own, within the larger Stockholm group, and that is just how they appear in Fig. 5.

Robinson’s study of the manuscripts St and Ra, summarized in Fig. 1, revealed their fundamental importance in the evolution of the Svipdagsmál tradition. Some two-thirds of all the manuscripts, thirty-one of the total forty-six, appear to have descended either from St or Ra, or a manuscript (the hypothetical X3) very close to Ra. Thus, although St and Ra are very similar in absolute terms, in evolutionary terms they are far apart: they represent the twin roots from which most of the manuscripts derive. The cladistic analysis manages to show this: Fig. 5 places five nodes between St and Ra, and from these five nodes all the other manuscripts descend.

In addition to producing an estimate of the stemma, PAUP can also produce a list of all the derived character states—the variant readings—arranged according to the branches of the tree. For example, the reading doetra in Gróugaldr 1 occurs in all nine B manuscripts. Accordingly, PAUP’s list of derived states has this reading being introduced at the root of the B manuscript branch. Thus, one can use PAUP to discover just what readings are characteristic of what groups of manuscripts. The utility of this for a textual critic is obvious: compare the rather clumsy database queries Robinson (1989b) used in his previous analysis of the Svipdagsmál manuscripts, or those used by Xhardez (1994) in his work on some fifty manuscripts of a twelfth century text.

6. Limitations

Although the analyses described here may seem substantial, nearly duplicating in a few minutes what it had taken several months to accomplish earlier, they are in fact very preliminary. Had the data been initially recorded with a view to their subsequent use in cladistic analysis, more information about the relations of different readings to one another could have been incorporated into the analysis, and the inclusion of this information would almost certainly have improved the result. It could have been specified in advance, for example, that certain readings represented omissions which were unlikely to be restored, or that other readings almost certainly arose in a particular sequence, and this information could have been taken into account by PAUP as it evaluated the fit of the data to different trees. The inclusion of such considerations is standard in cladistic analyses in evolutionary biology. There is some danger in including transformation assumptions of this sort in an initial analysis, of course, because it is possible to adjust the data put into the program to such an extent that whatever result one desires can be made to come out. Nevertheless, judicious incorporation of such assumptions, as long as they are clearly stated, is a reasonable procedure, and their incorporation would have improved PAUP’s estimate of the true history of the Svipdagsmál tradition.

Even if certain initial assumptions had been added to the analysis, however, there may still have been some limitations on the result. The greatest difficulties for the cladistic approach lie in the areas of manuscript contamination. Cladistic analysis effectively assumes that instances of vertical transmission will outnumber instances of horizontal transmission. This is broadly true of the mass of variants in most manuscript traditions, hence our general success with the Svipdagsmál material. But there may be subgroups of variants in subgroups of manuscripts that have been much influenced by horizontal transmission. There are, for example, a large number of variants found as marginalia in several groups of Svipdagsmál manuscripts which appear to have been borrowed from the text of other distinct groups, and the inclusion of these variants led to some deformation in the tree produced by the cladistic analysis. As a case in point, because of large scale contamination of the Langebek manuscripts by B manuscript readings, the Langebek manuscripts appear far closer to the B manuscripts in Fig. 5 than they should. This incorrect placement of the Langebek manuscripts had other, potentially serious, consequences. Robinson’s analysis of the manuscripts suggested that the B group had descended from Ra, or a manuscript very close to Ra, probably via manuscript O. The evidence for this is a set of some twenty-six readings found in Ra, also in O, and thence characteristically in the B manuscripts (Robinson 1991, 207). In order to accommodate the Langebek manuscripts (none of which have any of these twenty-six readings) somewhere between Ra and O in the cladogram, PAUP had to suppose that these twenty-six readings were first removed along the branch marked a (hence their absence from the manuscripts below that point, including the Langebek group), and then restored along the branch marked b (hence their presence in the manuscripts below that point, including O and the B manuscripts). This obscures the most likely flow of readings and makes Ra, O, and the B manuscripts appear rather more distantly related than they actually are.

There is a similar problem with manuscript J and the three manuscripts Gu, 11, and 682. There is strong external evidence for the close relationship of Gu, 11, and 682 (Robinson 1991, 427–28, 432–34; Fig. 1), with 11 and 682 both being copies of Gu, and Gu itself being a copy of St. Accordingly, the cladistic analysis succeeds in placing these three manuscripts directly adjacent to one another in Fig. 5. Robinson’s account of the manuscripts has J descending quite separately from either St or Ra, and appearing among a group of manuscripts at the centre of the table which lack any strong affiliations with any other manuscripts. But the cladistic analysis has grouped J with the three manuscripts Gu, 11, and 682; not only this, it has moved Gu, 11, and 682 much farther away from their direct ancestor St than is correct. Examination of the variants that PAUP judged to have been introduced along the branch marked c in Fig. 5 revealed the reason for this error. In the second poem of the Svipdagsmál sequence a question formula is repeated eighteen times over. Most manuscripts abbreviate this formula in one way or other, some giving the initial letter of each word, some just giving the first one or two words, and so on. Four manuscripts alone spell out every word of the whole question sequence on each repetition: they are the four manuscripts J, Gu, 11, and 682. Clearly, 11 and 682 have simply inherited this from Gu. Clearly too, in view of the lack of any other evidence linking J and the three manuscripts Gu, 11, and 682—J has only six of the twenty-eight variants which characterize the manuscripts descended from St, while Gu, 11, and 682 have respectively twenty-five, sixteen, and twenty-five (Robinson 1991, 178, 221)—it is simple accident that the scribe of J happened to spell out every instance of the formula just as the scribes of the other three manuscripts did. But this accident has caused the group Gu, 11, and 682 to be placed next to J and much further away from their direct ancestor St than is correct. Once more, this distorts the flow of readings: it requires us to suppose that most of the St variants present in Gu, 11, and 682 were removed before point c, and then restored along the branch marked d.

A further difficulty with the use of cladistics programs with manuscript data is that these programs tend to produce bifid trees: in Fig. 5, for example, every branch that divides always divides in two. Textual critics who recall Joseph Bédier’s scathing denunciation of the tendency of textual editors to create bifid stemmata and only bifid stemmata will be dubious. It is simply not true that in the history of a manuscript tradition each exemplar is copied twice and just twice. M.D. Reeve reports eighteen textual traditions of classical Latin texts in which an archetype and at least two descendants survive. In six of these, he finds stemmata with more than two branches and no certain cases of bipartite stemmata among the the other twelve (1986, 60). In the Svipdagsmál tradition there appear to have been at least three separate copyings of each of the manuscripts Gu and L, and the Svipdagsmál stemma as a whole grows from three distinct basal branches (Fig. 1). In creating the bifid tree shown in Fig. 5, PAUP was interpreting the data in a very strict sense. If three independent copies are made from a single ancestral manuscript, unless all three copies agree with each other in exactly the same number of introduced readings not present in their common ancestor, strict interpretation of the data will force the conclusion that two of these manuscripts are more closely related to one another (have a greater number of introduced readings or derived states in common) than either is to the third manuscript. Mere chance will see to it that some two of the three manuscripts will agree on a greater number of introduced readings than will either of these two with the third. In these circumstances, strict cladistic analysis will presume the existence of an intermediate ancestor for the two manuscripts sharing the greatest number of introduced readings. Thus, for three manuscripts A, B, and C all copied from a single ancestor X (Fig. 6, left), but with A and B having by chance a greater number of introduced readings in common than either has with C, strict cladistic analysis will generate the tree shown on the right in Fig. 6, hypothesizing a hyparchetype X1 as the ancestor of A and B but not of C. The textual critic must decide when this procedure is justified and when it is not. It should be clearly understood that we are not arguing that the application of numerical cladistic techniques obviates the need for critical thought.8 Quite the contrary in fact: numerical cladistic analyses can provide estimates of manuscript histories very quickly, so that thought may be applied to the details of those histories with greater efficiency.

7. Future Prospects

The limitations just described should not detract from the great value of the cladistic approach to manuscript studies. Had the results of this analysis been available four years ago they would have saved Robinson some six months’ work. Further, database techniques and traditional editorial scrutiny of variants would have made the errors in the analysis relatively easy to uncover. The value of numerical cladistic analyses is that they can help scholars to see where to look. Preliminary cladistic analyses of several other manuscript traditions9 up to the time of this writing have confirmed the promise of the approach, and have revealed several cases of relations which previous scholars had not suggested, but which even brief examination shows might be well-founded. These cases are now under study. The possible permutations of manuscripts in even a small tradition are so numerous that a scholar might easily miss the evidence for an important manuscript relationship simply because he or she never thought to make a particular comparison. Cladistics can generate a road map of a manuscript tradition, and this road map can be used to guide further research, either by database analysis of groups of variants, or by traditional scholarly assessment of key variants. Other analytical tools that are in use in the systematics community will be valuable in these further investigations as well, most notably the computer program MacClade (Maddison and Maddison 1989) which permits interactive investigation of trees and character state distributions.

Cladistic analysis works with manuscripts, and with due care, the evolution of many complex manuscript traditions can now be reconstructed. This may have considerable implications for three areas of study: textual editing, linguistic analysis, and cultural history. For textual editors, the recovery of the history of important and complex traditions, such as those of the Canterbury Tales or Piers Plowman, will change how these texts are edited. For students of language, it will become possible to analyze the changing linguistic forms found in manuscripts across entire textual traditions. For students of culture, studies of the reception of texts will become easier as the histories of those texts become more clearly established.

It may even be possible for some insights to pass from textual criticism back into systematics. One of the troublesome problems in systematics is the recognition and historical interpretation of hybrids—crosses between members of distinct branches of the evolutionary tree. This is quite similar to the problem of detecting contamination in manuscript traditions. If textual scholars apply themselves to the problem of manuscript contamination and its relation to cladistic analysis, they may be able to make a significant contribution to the historical study of hybridization in evolutionary biology as well. This would be a valuable contribution indeed to the study of ‘trees of history’ wherever they may be found.


For assistance and discussion we are grateful to A.R. Ives, G.C. Mayer, K. de Queiroz, E. Sober, D.L. Swofford, J.E. Wills, C.M. Sperberg-McQueen, P. James, M. Godden and the HUMANIST electronic discussion group. This study was supported in part by a Leverhulme Trust grant to PMWR, and by a Smithsonian Institution Postdoctoral Fellowship and a U.S. National Science Foundation Postdoctoral Fellowship (DIR-9103325) to RJO.


1. The other three manuscripts were not transcribed in full into electronic form, and were not incorporated into the analyses reported here.

2. Gösta Holm (Saga och sed 1972, 48–80) has described a stemma published by Carl Johan Schlyter in 1827, four years earlier than the Zumpt stemma we reproduced above.

3. It is interesting to note that evolutionary trees traditionally point upwards, whereas manuscript stemmata traditionally point downwards (Cameron 1987). We have elected as a compromise to draw our own trees (Figs. 4 and 5) pointing to the right. See O’Hara (1991, 1992) for a discussion of directionality in representations of evolutionary history.

4. ‘Die Feststellung eines Textes nach Ueberlieferung ist eine streng historische Arbeit’.

5. The number of possible trees for a given number of endpoints has been calculated by Felsenstein (1977) in systematics, and in parallel by Flight (1990) in stemmatics.

6. Our collaboration on this project began in July 1991 when Robinson, a textual scholar, posted a challenge to the HUMANIST electronic discussion group to discover whether anyone could duplicate his Svipdagsmál stemma using only the raw matrix of agreements and disagreements among the manuscripts. O’Hara, an evolutionary biologist, responded.

7. This tree was estimated from a matrix of 43 manuscripts by 3138 readings, of which 2063 were informative. (A copy of the complete matrix is available from the authors upon request.) It is the single shortest tree found by PAUP’s heuristic search procedure under 500 random permutations of the addition sequence, and has a length of 8249, a consistency index of 0.38, and a retention index of 0.625 (excluding uninformative characters l=7181 and c.i.=0.287). It is important to understand that this tree is, in the terminology of cladistics, an ‘unrooted tree’. This means that no information was included in the analysis to specify which readings were ancestral and which were derived, and that the tree can theoretically be ‘read’ outward from any starting point. We have rooted the tree arbitrarily so as to make it correspond approximately with the rooting of Fig. 1. The tree could have been rooted in a non-arbitrary manner by adding to the data matrix an additional row that specified as many of the ancestral readings as could be determined by traditional principles of textual criticism such as lectio difficilior.

8. Researchers in systematics who have produced superficial numerical analyses of cladistic data have themselves been vigorously criticized. See for example the recent commentaries on the work of Cann et al. (1987) and Vigilant et al. (1991) by Maddison (1991), Hedges et al. (1992), and Templeton (1992).

9. The twenty-two manuscripts of Dante’s Monarchia, the sixty-five of the Old Norse Song of the Sun, twenty-seven of Redaction IV of the Visio Pauli, forty-two of part of Chaucer’s ‘Wife of Bath’s Prologue’, eight editions of a Hume essay, and six manuscripts of the B version of Langland’s Piers Plowman.


{Figure Captions}

Fig. 1. Relationships of the Svipdagsmál manuscripts, after Robinson (1991)

Note: Branch lengths and branching angles are arbitrary, and branches may be rotated about nodes arbitrarily. Arrows indicate relationships confirmed by external evidence. X–X9 are hypothetical ancestors. Ra may be identical with X3 rather than a copy of it, and He may be either a copy of X5 or identical with X5.

Fig. 2. Diagram of evolutionary descent, from Darwin (1859)

Note: This was the only illustration in Darwin’s Origin of Species, and it initiated a long period of tree-making in systematic biology.

Fig. 3. The ‘stemma codicum’ of the manuscripts of Cicero’s Verrine Orations, published by C.G. Zumpt (1831, xxxviii)2

Note: The manuscripts are divided into two branches, descending from ‘Cod. antiquus deperditus’ and ‘Cod. Vatican. rescriptus’.

Fig. 4. Elementary cladistic analysis

Note: On the left is a data matrix of five taxa by three characters; 0 stands for the ancestral state in each character. Tick marks on the branches represent character state changes as the characters from the matrix are successively added to the tree (left to right). Branch lengths are arbitrary, and branches may be rotated about nodes arbitrarily.

Fig. 5. Estimate of the history of the Svipdagsmál manuscripts generated by the cladistics program PAUP

Note: Some of the major groupings of manuscripts common to this tree and to Robinson’s stemma (Fig. 1) are indicated. Horizontal branch lengths are proportional to the number of character state changes along each branch. Vertical branch lengths are arbitrary, and branches may be rotated about nodes arbitrarily. See Note 6 for additional details.

Fig. 6. Postulating hypothetical ancestors in cladistic analysis

Note: Let the tree on the left represent the true history of three manuscripts (A, B, and C), all copied from the same ancestor (X). If by chance A and B happen to share a greater number of derived readings with each other than either does with C, strict application of cladistic procedures will lead to the conclusion that A and B had an ancestor in common (X1) that they did not share with C, an ancestor that exhibited the set of derived readings common to A and B.

{Author Note}

Peter M.W. Robinson is Research Officer of the Computers and Manuscripts Project within Oxford University Computing Services, has edited Old Norse poetic texts, and is head of the Textual Criticism working group of the Text Encoding Initiative.

Robert J. O’Hara is U.S. National Science Foundation Postdoctoral Fellow in philosophy and evolution at the University of Wisconsin–Madison, and Adjunct Curator in the University of Wisconsin Zoological Museum.

© RJO 1995–2022