Friday, September 4, 2015
On what grounds can we trust a theory in the absence of empirical confirmation?
Thanks, Lubos, for your kind invitation to write a guest blog on non-empirical theory confirmation (which I have recently presented in the book String Theory and the Scientific Method, CUP 2013). As a long-time follower of this blog (who, I may dare to add, fervently disagrees with much of its non-physical content) I am very glad to do so. Fundamental physics today faces an unusual situation. Virtually all fundamental theories that have been devised during the last four decades still lack conclusive empirical confirmation. While the details with respect to empirical support and prospects for conclusive empirical testing vary from case to case, this general situation is exemplified by theories like low energy supersymmetry, grand unified theories, cosmic inflation or string theory. The fact that physics is characterised by decades of continuous work on empirically unconfirmed theories turns the non-empirical assessment of those theories' chances of being viable into an important element of the scientific process. Despite the scarcity of empirical support, many physicists working on the above-mentioned theories have developed substantial trust in their theories' viability based on an overall assessment of the physical context and the theories' qualities. In particular in the cases of string theory and cosmic inflation, that trust has been harshly criticised by others as unjustified and incompatible with basic principles of scientific reasoning. The critics argue that empirical confirmation is the only possible scientific basis for holding a theory viable. Relying on other considerations in their eyes amounts to abandoning necessary scientific restraint and leads to a relapse into pre-scientific modes of reasoning. The crux of the argument is the concept of scientific theory confirmation. In my recent book, I argue that the critics' wholesale condemnation of non-empirical reasons for having trust in a theory's viability is caused by their adherence to an oversimplified and inadequate understanding of scientific confirmation that, unfortunately, has dominated the philosophy of science in the 20th century. I propose an understanding of theory confirmation that is broader than what is commonly understood as empirical confirmation and therefore does provide a basis for acknowledging the general soundness of the lines of reasoning that lead physicists towards trusting their theories on a non-empirical basis. (To be sure, this does not imply that non-empirical arguments for a theory's viability are always convincing. It just means that arguments of that kind can be convincing in principle.) The canonical understanding of scientific confirmation, which can for example be found in classical hypothetico-deductivism and in most presentations of Bayesian confirmation theory, is the following. A theory can be confirmed only by empirical data that is predicted by that theory. Agreement between prediction and data amounts to confirmation, disagreement amounts to disconfirmation. Now, we may take this as a primitive definition of confirmation, in which case it makes no sense to question it. The problem is, however, that most scientists have a wider intuitive understanding of confirmation. According to that wider understanding, scientific confirmation is the scientifically supported generation of trust in a theory's viability based on observation. That wider intuitive understanding is implicitly adopted by the critics of non-empirical reasons for trusting a theory: they assume that, in the absence of confirmation, there can't be any good reasons for trusting a theory. My point of departure thus is the understanding that our concept of confirmation should cover all observation-based scientifically supported reasons for believing in a theory's viability. If so, however, it is by no means clear that the observations involved must always be predicted by the given theory. In fact, we can find cases of scientific reasoning where that is quite obviously not the case. A striking example is the Higgs hypothesis. High energy physicists were highly confident that some kind of Higgs particle (whatever the details) existed long before a Higgs particle was discovered in 2012. Their confidence was based on an assessment of the scientific context and their overall experience with predictive success in physics. Even before 2012, it would have been difficult to deny the scientific legitimacy of that assessment. It would be even more implausible today, after that assessment has been vindicated by the LHC. Clearly, there is an important difference between the status of the Higgs hypothesis before and after its successful empirical testing in 2011/2012. That difference can be upheld by distinguishing two different kinds of confirmation. Empirical confirmation is based on the empirical testing of the theory's predictions. Non-empirical confirmation is based on observations that are not of the kind that can be predicted by the theory. Conclusive empirical confirmation is more powerful than non-empirical prediction. But non-empirical confirmation can also provide strong reasons for believing in a theory's viability. What are the observations that generate non-empirical confirmation in physics today? Three main kinds of argument, each relying on one type of observation can be found when looking at the research process. They don't work in isolation but only acquire strength in conjunction. The first and most straightforward argument is the no alternatives argument (NAA). Physicists infer the probable viability of a theory that solves a specific physical problem from the observation that, despite extensive efforts to do so, no alternative theory that solves this problem has been found. Trust in the Higgs hypothesis before empirical confirmation was crucially based on the fact that the Higgs hypothesis was the only known convincing theory for generating the observed mass spectrum within the empirically well-confirmed framework of gauge field theory. In the same vein, trust in string theory is based on the understanding that there is no other known approach for a coherent theory of all fundamental interactions. On its own, NAA has one obvious weakness: scientists might just have not been clever enough to find the alternatives that do exist. In order to take NAA seriously, one therefore needs a method of assessing whether or not scientists in the field typically are capable of finding the viable theories. The argument of meta-inductive inference from predictive success in the research field (MIA) does the job. Scientists observe that in similar contexts, theories without known alternatives turned out to be successful once empirically tested. Both, pre-discovery trust the Higgs hypothsis and today's trust in string theory gain strength from the observation that standard model predictions were highly successful empirically. One important caveat remains, however. It often seems questionable whether previous examples of predictive success and the new theory under scrutiny are sufficiently similar to justify the use of MIA. In some cases, for example in the Higgs case, the concept under scrutiny and previous examples of predictive success are so closely related to each other that the deployment of MIA looks fairly unproblematic. NAA and MIA in conjunction thus were sufficient in the Higgs case for generating a high degree of trust in the theory. In other cases, like string theory, the comparison with earlier cases of predictive success is more contentious. In many respects, string theory does constitute a direct continuation of the high energy physics research program that was so successful in the case of the standard model. But its evolution differs substantially from that of its predecessors. The far higher level of complexity of the mathematical problems involved makes it far more difficult to approach a complete theory. This higher level of complexity may throw the justification for a deployment of MIA into doubt. In cases like that, it is crucial to have a third argument indicating that, despite the high complexity of the theory in question, scientists are still capable of finding their way through the 'conceptual labyrinth' they face. The argument that can be used to that end is the argument from unexpected explanatory interconnections (UEA). The observation on which UEA is based is the following: scientists develop a theory in order to solve a specific theory. Later it turns out that this theory also solves other conceptual problems it was not developed to solve. This is taken as an indicator of the theory's viability. UEA is the theory-based 'cousin' of the well known data-based argument of novel predictive success. The latter relies on the observation that a theory that was developed based on a given set of empirical data correctly predicts new data that had not entered the process of theory construction. UEA now replaces novel empirical prediction by unexpected explanation. The most well-known example of UEA in the case of string theory is its role in understanding black hole entropy. String theory was proposed as a universal theory of all interactions because it was understood to imply the existence of a graviton and suspected to be capable of avoiding the problem of non-renormalizability faced by field theoretical approaches to quantum gravity. Closer investigations of the theory's structure later revealed that - at least in special cases - it allowed for the exact derivation of the known macro-physical black hole entropy law from micro-physical stringy structure. Consideration about black hole entropy, however, had not entered the construction of string theory. Beyond this particular example, string theory offers a considerable number of other unexpected explanatory interconnections that allow for the deployment of UEA. String theorists asked to what extent the think NAA, MIA and UEA influence their trust in their theory often answer that, while NAA and MIA do provide necessary stepping stones for trust in the theory, it is UEA that is the crucial reason for trusting the theory. NAA, MIA and UEA are applicable in a wide range of cases in physics. Their deployment is by no means confined to empirically unconfirmed theories. NAA and MIA play a very important role in understanding the significance of empirical theory confirmation. The continuity between non-empirical confirmation and the assessment of empirical confirmation based on NAA and MIA can be seen nicely by having another look at the example of the Higgs discovery. As argued above, the Higgs hypothesis was believed before 2012 based on NAA and MIA. But only the empirical discovery of a Higgs particle implied that calculations of the background for future scattering experiments had to contain Higgs contributions. That implication is based on the fact that the discovery of a particle in a specific experimental context is taken to be a reliable basis for having trust in that particle's further empirical implications. But why is that so? It relies on the very same types of consideration that had generated trust in the Higgs hypothesis already prior to discovery. First, no alternative theoretical conception is available that can account for the measured signal without having those further empirical implications (NAA). And second, in comparable cases of particle discoveries in the past trust in the particle's further empirical implications was mostly vindicated by further experimentation (MIA). Non-empirical confirmation in this light is no new mode of reasoning in physics. Very similar lines of reasoning have played a perfectly respectable role in the assessment of the conceptual significance of empirical confirmation throughout the 20th century. What has changed is the perceived power of non-empirical considerations already prior to empirical testing of the theory. While NAA, MIA and UEA are firmly rooted in the history of physical reasoning, string theory does add one entirely new kind of argument that can contribute to the strength of non-empirical confirmation. String theory contains a final theory claim, i.e. the claim that, if string theory is a viable theory at its own characteristic scale, it won't ever have to be superseded by an empirically distinguishable new theory. The future of theoretical conceptualization in that case would be devoted to fully developing the theory from the basic posits that are already known rather than to searching for new basic posits that are emprically more adequate. Though the character of string theory's final theory claim is not easy to understand from a philosophical perspective, it may shed new light on the epistemic status of string theory. My self-imposed space constraints for this blog don't allow a more far-reaching discussion of this question. I just want to point out that final theory claims seem to constitute a very interesting new twist to the question of non-empirical confirmation. For the remainder of this text, though, I want to confine my analysis to the role of the three 'classical' arguments NAA, MIA and UEA. Let us first address an important point. In order to be convincing, theory confirmation must not be a one way street. If a certain type of observation has the potential to confirm a theory, it must also have the potential to dis-confirm it. Empirical confirmation trivially fulfils that condition. For any set of empirical data that agrees with a theory's prediction, there are many others that disagree with it and therefore, if actually measured, would dis-confirm the theory. NAA, MIA and UEA fulfil that condition as well. The observation that no alternatives to a theory have been found could, in principle, always be overridden by future observations that scientists do find alternatives later on. That later observation would reduce the trust in the initial theory and therefore amount to that theory's non-empirical dis-confirmation. Likewise, an observed trend of predictive success in a research field could later be overridden by a series of instances where a theory that was well trusted on non-empirical grounds turned out to disagree with empirical tests once they became possible. In the case of UEA, the observation that no unexpected explanatory interconnections show up would be taken to speak against a theory's viability. And once unexpected interconnections have been found, it could still happen that a more careful conceptual analysis reveals them to be the result of elementary structural characteristics of theory building in the given context that are not confined to the specific theory in question. To conclude, the three non-empirical arguments are not structurally biased in favour of confirmation but may just as well provide indications against a theory's viability. Next, I briefly want to touch a more philosophical level of analysis. Empirical confirmation is based on a prediction of the confirmed theory that agrees with an observation. In the case of non-empirical confirmation, to the contrary, the confirming observations are not predicted by the theory. How can one understand the mechanism that makes those observations confirm the theory? It turns out that an element of successful prediction is involved in non-empirical confirmation as well. That element, however, is placed at the meta-level of understanding the context of theory building. More specifically, the claim that is tested at the meta-level is a claim on the spectrum of possible scientific alternatives to the known theory. The observations on which NAA, MIA and UEA rely are all predicted by the meta-level hypothesis that the spectrum of possible scientific alternatives to the theory in question is very limited. Inversely, NAA, MIA and UEA indicate that the spectrum of unconceived alternatives to the known theory is strongly limited. Let us, for the sake of simplicity, just consider the most extreme form of this meta-level hypothesis, namely the hypothesis that, in all research contexts in the scientific field, there are no possible alternatives to the viable theory at all. This radical hypothesis predicts 1: that no alternatives are found because there aren't any (NAA); 2: that a theory that has been developed will always be predictively successful, given that there exists a predictively successful theory at all (MIA); and 3: that that a theory that has been developed for one specific reason will explain all other things as well, because there are no alternatives that could (UEA). In order to use such claims of "limitations to scientific underdetermination", as I call them, as a serious foundation for non-empirical confirmation, one would have to say more on the criteria for accepting a theory as scientific, on how to individuate theories, etc. In this presentation, it shall suffice to give the general flavour of the line of reasoning: non-empirical confirmation is a natural extension of empirical confirmation that places the agreement between observation and the prediction of a hypothesis at the meta-level of theory dynamics. A clearer understanding of the mechanism of non-empirical confirmation and its close relation to empirical confirmation can be acquired based on a formalization of the arguments within the framework of Bayesian confirmation theory. An analysis of this kind has been carried out for NAA (which is the simplest case) in "The No Alternatives Argument", Dawid, Hartmann and Sprenger BJPS 66(1), 213-34, 2015. A number of worries have been raised with respect to the concept of non-empirical confirmation. Let me, in the last part of this text, address a few of them. It has been argued (e.g. by Sabine Hossenfelder) that arguments of non-empirical confirmation are sociological and therefore don't constitute proper scientific reasoning. This claim may be read in two different ways. In its radical form, it would amount to the statement that there is no factual scientific basis to non-empirical confirmation at all. Confidence in a theory on that account would be driven entirely by sociological mechanisms in the physics community and only be camouflaged ex post by fake rational reasoning. The present text in its entirety aims at demonstrating that such an understanding of non-empirical confirmation is highly inadequate. A more moderate reading of the sociology claim is the following: there may be a factual core to non-empirical confirmation, but it is so difficult to disentangle from sociological factors that science is better off when fully discarding non-empirical confirmation. I concede that the role of sociology is trickier with respect to deployments of non-empirical confirmation than in cases where conclusive empirical confirmation is to be had. But I would argue that it is must always be the aim of good science to extract all factual information that is provided by an investigation. If the existence of a sociological element in scientific analysis would justify discarding that analysis, quite some empirical data analysis had to be discarded as well. To give a recent example: the year 2015 witnessed considerable differences of opinion among physicists interpreting the empirical data collected by BICEP2, which might be understood to a certain degree by sociological factors involved. No-one would have suggested to discard the debate on the interpretation of that data as scientifically worthless on those grounds. I suggest that the very same point of view should also be taken with respect to non-empirical confirmation. It has also been suggested (e.g. by George Ellis and Joseph Silk) that non-empirical confirmation may lead to a disregard for empirical data and therefore to the abandonment of a pivotal principle of scientific reasoning. This worry is based on a misreading of non-empirical confirmation. Accepting the importance of non-empirical confirmation by no means devaluates the search for empirical confirmation. To the contrary, empirical confirmation is crucial for the functioning of non-empirical confirmation in at least two ways. Firstly, non-empirical confirmation indicates the viability of a theory. But a theory's viability is defined as: the theory's empirical predictions would turn out correct if they could be specified and empirically tested. Conclusive empirical confirmation therefore remains the ultimate judge of a theory's viability - and thus the ultimate goal of science. Secondly MIA, one cornerstone of non-empirical confirmation, relies on empirical confirmation elsewhere in the research field. Therefore, if empirical confirmation ended in the entire research field, that would remove the possibility of testing non-empirical confirmation strategies and, in the long run, make them dysfunctional. Non-empirical confirmation itself thus highlights the importance of testing theories empirically whenever possible. It implies, though, that not having empirical confirmation must not be equated with knowing nothing about the theory's chances of being viable. Finally, it has been argued (e.g. by Lee Smolin) that non-empirical confirmation further strengthens the dominant research program and therefore in an unhealthy way contributes to thinning out the search for alternatives perspectives that may turn out productive later on. To a given extent, that is correct. Taking non-empirical confirmation seriously does support the focus on those research strategies that generate theories with a considerable degree of non-empirical confirmation. I would argue, however, that this is, by and large, a positive effect. It is an important element of successful science to understand which approaches merit further investigations and which don't. But a very important second point must be added. The way non-empirical confirmation has been presented, it is a technique for understanding the spectrum of possible alternatives to the theory one knows. One crucial test in that respect is to check whether serious and extensive search for alternatives has produced any alternative theories (This is the basis for NAA). Therefore, the search for alternatives is a crucial element of non-empirical confirmation. Non-empirical confirmation does exactly the opposite from denying the value of the search for alternatives: it adds a new reason why it is important. The search for alternatives remains fruitful even if the alternative strands of research fail to produce coherent theories. In that case the understanding that none of the alternative approaches has succeeded gives an important contribution to the non-empirical confirmation of the theory that is available. So what is the status of non-empirical confirmation? The arguments I present support the general relevance of non-empirical confirmation in physics. In the absence of empirical confirmation, non-empirical confirmation can provide a strong case for taking a theory to be viable. This does by no means render empirical confirmation obsolete. Conclusive empirical testing will always trump non-empirical confirmation and therefore remains the ultimate goal in science. Arguments of non-empirical confirmation can in some cases lead to a nearly consensual assessment in the physics community (see the trust in the Higgs particle before 2012). In other cases, they can also be more controversial. As in all contexts of scientific inquiry, arguments presented can be balanced and well founded but may also be exaggerated and unsound in some cases. The actual strength of each specific case of non-empirical confirmation has to be assessed and discussed by the physicists concerned with the given theory based on a careful scientific analysis of the particular case. Criticism of cases on non-empirical confirmation at that level constitutes an integral part of theory assessment. I suggest, however, that a whole-sale verdict that non-empirical theory confirmation is unscientific and should not be taken seriously does not do justice to the actual research process in physics and obscures the actual state of contemporary physics by disregarding one important element of scientific analysis. Richard Dawid Center for Mathematical Philosophy LMU Munich