Abstract
Striving for a probabilistic explication of coherence, scholars proposed a distinction between agreement and striking agreement. In this paper I argue that only the former should be considered a genuine concept of coherence. In a second step the relation between coherence and reliability is assessed. I show that it is possible to concur with common intuitions regarding the impact of coherence on reliability in various types of witness scenarios by means of an agreement measure of coherence. Highlighting the need to separate the impact of coherence and specificity on reliability it is finally shown that a recently proposed vindication of the Shogenji measure qua measure of coherence vanishes.
This is a preview of subscription content, log in to check access.
Notes
 1.
 2.
 3.
 4.
 5.
 6.
 7.
Here and in what follows I assume the probability measure to be regular so that \(\Pr (A)=0\) only if \(A\) is selfcontradictory.
 8.
 9.
This example is proposed as a test case for coherence measures by Schippers and Siebel (2014).
 10.
For a proof of this assertion see Appendix 1.
 11.
 12.
This latter property, the “BarHillelCarnap semantic paradox” (Floridi (2004), p. 198) , might seem curious at first sight. However, (iii) follows naturally from the assumption that \(A\)’s information content is related to the amount of state descriptions precluded by \(A\). A tautology precludes no state description at all and is consequently assigned the lowest possible degree of informativity. On the other hand, a contradiction precludes every possible state description and is accordingly maximally informative. In this sense, BarHillel & Carnap state that “a false sentence which happens to say much is thereby highly informative in our sense. [\(\ldots \)]. A selfcontradictory sentence asserts too much; it is too informative to be true” (p. 229).
 13.
Note that this proposal matches Shogenji’s (1999) definition of the “total individual strength” of a set of propositions, calculated as the product of the prior probabilities of each member of the set. It can easily be shown that the total individual strength of set \(S_1\) exceeds the one of \(S_2\) iff \(\sigma (S_1)>\sigma (S_2)\) (proof omitted).
 14.
Cf. BonJour (1985), p. 148.
 15.
 16.
A proof is given in Appendix 2.
 17.
For example Schubert assumes that \(\Pr (A\,\,\langle A^i\rangle ,\,\lnot R_i)=\Pr (A\,\,\lnot R_i)\). However, given , it is easily shown that \(\Pr (A\,\,\lnot R_i)=\alpha \).
 18.
A proof is given in Schubert (2012a).
 19.
A proof is given in Appendix 3.
 20.
A proof is given in Appendix 4.
 21.
This measure is proposed in Shogenji (1999).
 22.
Further deficiencies of Shogenji’s measure are spotted in Akiba (2000); Fitelson (2003); Glass (2005); Koscholke (2013); Schupbach (2011); Siebel (2005); Wheeler (2009). Positive findings regarding the ability of \({{\mathcal C}_{\mathcal A}}\) to cope with various test cases for probabilistic measures of coherence can be found in Koscholke (2013) and Roche (2013).
 23.
 24.
The distributions underlying these results are the ones given above in the corresponding proof for the Shogenji measure.
 25.
I am grateful to an anonymous referee for spotting an error in the derivation.
References
Akiba, K. (2000). Shogenji’s probabilistic measure of coherence is incoherent. Analysis, 60, 356–359.
Angere, S. (2008). Coherence as a heuristic. Mind, 117, 1–26.
BarHillel, Y., & Carnap, R. (1953). Semantic information. British Journal for the Philosophy of Science, 4, 147–157.
BonJour, L. (1985). The structure of empirical knowledge. Cambridge, London: Harvard University Press.
Bovens, L., & Hartmann, S. (2003). Bayesian epistemology. Oxford: Oxford University Press.
Bovens, L., & Olsson, E. J. (2000). Coherentism, reliability and Bayesian networks. Mind, 109, 685–719.
Brössel, P. (2013). The problem of measure sensitivity redux. Philosophy of Science, 80, 378–397.
Carnap, R. (1962). The logical foundations of probability (2nd ed.). Chicago: The University of Chicago Press.
Crupi, V., Chater, N., & Tentori, K. (2013). New axioms for probability and likelihood ratio measures. British Journal for the Philosophy of Science, 64, 189–204.
Crupi, V., & Tentori, K. (2013). Confirmation as partial entailment: A representation theorem in inductive logic. Journal of Applied Logic, 11, 364–372.
Dietrich, F., & Moretti, L. (2005). On coherent sets and the transmission of confirmation. Philosophy of Science, 72, 403–424.
Douven, I., & Meijs, W. (2007). Measuring coherence. Synthese, 156, 405–425.
Fitelson, B. (1999). The plurality of Bayesian measures of confirmation and the problem of measure sensitivity. Philosophy of Science, 66, 362–378.
Fitelson, B. (2001). Studies in Bayesian confirmation theory. PhD thesis. University of WisonsinMadison.
Fitelson, B. (2003). A probabilistic theory of coherence. Analysis, 63, 194–199.
Fitelson, B. (2004). Two technical corrections to my coherence measure. http://fitelson.org/coherence2.pdf.
Floridi, L. (2004). Outline of a theory of strongly semantic information. Minds and Machines, 14, 197–221.
Glass, D. H. (2002). Coherence, explanation, and Bayesian networks. In M. O’Neill, et al. (Eds.), Artificial intelligence and cognitive science (pp. 177–182). Berlin, Heidelberg: Springer.
Glass, D. H. (2005). Problems with priors in probabilistic measures of coherence. Erkenntnis, 63, 375–385.
Glass, D. H. (2007). Coherence measures and inference to the best explanation. Synthese, 157, 275–296.
Hintikka, J. (1968). The varieties of information and scientific explanation. In B. van Rootselaar & J. E. Staal (Eds.), Logic, Methodology and Philosophy of Science, III: Proceedings of the 1967 International Congress (pp. 151–171). Amsterdam: NorthHolland.
Hintikka, J. (1970). On semantic information. In J. Hintikka & P. Suppes (Eds.), Information and inference (pp. 3–27). Dordrecht: Reidel.
Hintikka, J., & Pietarinen, J. (1966). Semantic information and inductive logic. In J. Hintikka & P. Suppes (Eds.), Aspects of inductive logic (pp. 96–112). Dordrecht: Reidel.
Koscholke, J. (2013). Last measure standing. Evaluating test cases for probabilistic coherence measures. Unpublished manuscript.
Lewis, C. I. (1946). An analysis of knowledge and valuation. Lasalle: Open Court.
Meijs, W. (2006). Coherence as generalized logical equivalence. Erkenntnis, 64, 231–252.
Meijs, W., & Douven, I. (2007). On the alleged impossibility of coherence. Synthese, 157, 347–360.
Moretti, L., & Akiba, K. (2007). Probabilistic measures of coherence and the problem of belief individuation. Synthese, 154, 73–95.
Olsson, E. J. (2002). What is the problem of coherence and truth? The Journal of Philosophy, 99, 246–272.
Olsson, E. J. (2005). Against coherence: Truth, probability and justification. Oxford: Oxford University Press.
Olsson, E. J., & Schubert, S. (2007). Reliability conducive measures of coherence. Synthese, 157, 297–308.
Pearl, J. (2000). Causality: Models, reasoning, and inference. New York: Cambridge University Press.
Roche, W. (2013). Coherence and probability. A probabilistic account of coherenc. In M. Araszkiewicz & J. Savelka (Eds.), Coherence: Insights from philosophy, jurisprudence and artificial intelligence (pp. 59–91). Dordrecht: Springer.
Schippers, M., & Siebel, M. (2012). Reassessing probabilistic measures of coherence. Unpublished manuscript.
Schippers, M., & Siebel, M. (2014). Inconsistent testimonies as a touchstone for coherence measures. Unpublished manuscript.
Schubert, S. (2011). Coherence and reliability: The case of overlapping testimonies. Erkenntnis, 74, 263–275.
Schubert, S. (2012a). Coherence reasoning and reliability: A defense of the Shogenji Measure. Synthese, 187, 305–319.
Schubert, S. (2012b). Is coherence conducive to reliability? Synthese, 187, 607–621.
Schupbach, J. N. (2008). On the alleged impossibility of Bayesian coherentism. Philosophical Studies, 141, 323–331.
Schupbach, J. N. (2011). New hope for Shogenji’s coherence measure. The British Journal for the Philosophy of Science, 62, 125–142.
Shannon, C. E. (1948). A mathematical theory of communication. Part I, Bell Systems Technical Journal, 27, 379–423.
Shannon, C. E., & Weaver, W. (1949). A mathematical model of communication. Urbana, IL: University of Illinois Press.
Shogenji, T. (1999). Is coherence truth conducive? Analysis, 59, 338–345.
Siebel, M. (2004). On Fitelson’s measure of coherence. Analysis, 64, 189–190.
Siebel, M. (2005). Against probabilistic measures of coherence. Erkenntnis, 63, 335–360.
Siebel, M., & Wolff, W. (2008). Equivalent testimonies as a touchstone for coherence measures. Synthese, 161, 167–182.
Wheeler, G. (2009). Focussed correlation and confirmation. The British Journal for the Philosophy of Science, 60, 79–100.
Wheeler, G. (2012). Explaining the limits of Olsson’s impossibility result. The Southern Journal of Philosophy, 50, 136–150.
Acknowledgments
I owe special thanks to Jakob Koscholke, Erik J. Olsson, Stefan Schubert, Mark Siebel and two anonymous reviewers for helpful comments or discussion. This work was supported by Grant SI 1731/11 to Mark Siebel from the Deutsche Forschungsgemeinschaft (DFG) as part of the priority program “New Frameworks of Rationality” (SPP 1516).
Author information
Appendices
1. Extant coherence measures violate \((\dagger )\)
Recall condition \((\dagger )\): consider a set \(S=\{A_1,\ldots ,A_n\}\) of propositions and two probability distributions \(\Pr \) and \(\Pr '\). If the posterior probability of each nonempty subset of propositions on each nonempty subset of the remaining propositions in the set for distribution \(\Pr \) exceed the corresponding posterior probabilities for \(\Pr '\), then \(S\) is more coherent on \(\Pr \) than on \(\Pr '\). In order to show that extant coherence measures violate this condition, I distinguish the following cases:
The Shogenji measure
For a set \(S=\{A_1,\ldots ,A_n\}\), the Shogenji measure reads as follows:
Let \(\{A,B\}\) be a set of propositions and consider the following probability distributions \(\Pr \) and \(\Pr '\): for \(\Pr \), \(\Pr (A\wedge B)=0.3\) and \(\Pr (A\wedge \lnot B)=0.2=\Pr (\lnot A\wedge B)\). For \(\Pr '\), \(\Pr '(A\wedge B)=0.525\) and \(\Pr '(A\wedge \lnot B)=\Pr '(\lnot A\wedge B)=0.225\). Then \(\Pr (A)=\Pr (B)=0.5\) and \(\Pr '(A)=\Pr '(B)=0.75\). Hence, \(\Pr (A\wedge B)=0.3>0.25=\Pr (A)\cdot \Pr (B)\) and \(\Pr '(A\wedge B)=0.525<0.5625=\Pr '(A)\cdot \Pr '(B)\). Thus, \({\mathcal C}_\mathrm{Sh}(A,B)\) is higher on probability distribution \(\Pr \) than on \(\Pr '\). On the other hand, \(\Pr (AB)=0.6<0.7\Pr '(AB)\) and also \(\Pr (BA)=0.6<0.7=\Pr '(BA)\). Thus, \({\mathcal C}_\mathrm{Sh}\) violates \((\dagger )\).
Mutual support measures
The basic idea underlying these measures is that coherent propositions mutually support each other. One way to cash out this idea is due to Douven and Meijs (2007). Their recipe for probabilistic measures of coherence averages the degree of mutual support for all pairs of nonempty, nonoverlapping subsets of the given set under consideration. In a nutshell, the idea is similar to the construction of the agreement measures \({{\mathcal C}_{\mathcal A}}\) except for the fact that Douven and Meijs suggest to implement probabilistic measures of incremental support. Basically, \(\mathfrak {s}\) is a measure of incremental support, if there is a threshold \(\beta \) such that for each pair of a hypothesis \(H\) and a piece of evidence \(E\),
Hence, evidence \(E\) is said to confirm \(H\) iff \(\mathfrak {s}(H,E)>\beta \), \(E\) is neutral towards \(H\) iff \(\mathfrak {s}(H,E)=\beta \), and \(E\) disconfirms \(H\) iff \(\mathfrak {s}(H,E)<\beta \). Now, given this qualitative explication of the term confirmation, it remains an open question whether two confirmatory pieces of evidence \(E_1\) and \(E_2\) are equally confirmatory or not. Similarly, given that \(E\) confirms both hypothesis \(H_1\) and \(H_2\) it will often be the case, intuitively, that one of the hypothesis is more confirmed than the other. By means of a probabilistic measure of confirmation, it is possible to answer these and related question: these measures are functions that assign to each hypothesisevidence pair \(H,E\) a number \(\alpha \) that represents the degree of confirmation that \(E\) confers upon \(H\). Obviously, there are many ways to spell out the notion of a degree of confirmation. Accordingly, a plethora of nonequivalent confirmation measures have been proposed.^{Footnote 23} Since all of these measures share property (11), we need not go into the details of the various measures. Instead, I will show that all measures that satisfy property (11) will inevitably violate \((\dagger )\).
Let \(\Pr \) and \(\Pr '\) be two probability distributions such that \(\Pr (AB)=\Pr (BA)=0.6\) and \(\Pr '(AB)=\Pr '(BA)=0.7\). Furthermore, assume that \(\Pr (A)=\Pr (B)=0.5\) and \(\Pr '(A)=\Pr '(B)=0.75\).^{Footnote 24} Now, on the one hand, given that \(\Pr (AB)<\Pr '(AB)\) and \(\Pr (BA)<\Pr '(BA)\), the set \(\{A,B\}\) is more coherent on probability distribution \(\Pr '\) than on \(\Pr \) according to \((\dagger )\). On the other hand, since \(\Pr (AB)>\Pr (A)\) and \(\Pr (BA)>\Pr (B)\), whereas \(\Pr '(AB)<\Pr '(A)\) and \(\Pr '(BA)<\Pr '(B)\), \(\{A,B\}\) must turn out less coherent on \(\Pr '\) than on \(\Pr \). This is due to the fact that for each confirmation measure \(\mathfrak {s}\), the \(\mathfrak {s}\)based coherence value for \(\{A,B\}\) is the straight average of \(\mathfrak {s}(A,B)\) and \(\mathfrak {s}(B,A)\). Since both these confirmation values are larger than \(\beta \) for \(\Pr \) and lower than \(\beta \) for \(\Pr '\), the corresponding coherence values must be so likewise. Hence, \(\{A,B\}\) is less coherence on \(\Pr '\) than on \(\Pr \) irrespective of the chosen confirmation measure \(\mathfrak {s}\).
So far there is only one (family of) coherence measure with an independent standing in the literature next to \({{\mathcal C}_{\mathcal A}}\) that might satisfy \((\dagger )\). These are the overlap measure of coherence, independently proposed by Glass (2002) and Olsson (2002), and it’s refined version endorsed by Meijs (2006). However, it is an open question whether these measures do indeed satisfy \((\dagger )\). Both these measures are not ordinally equivalent to \({{\mathcal C}_{\mathcal A}}\). Furthermore, both violate the additivityconstraint mentioned in Sect. 2.
2. Proof of Theorem 1
Theorem 1: In a scenario with two testimonies that satisfies conditions (4)(6) and fixed equal individual reliability the following comparative assertions are true:

(i)
Given fixed equal informativity \({\text {inf}}(A_1)={\text {inf}}(A_2)\), \({\Delta }(R_1)\) covaries with \({{\mathcal C}_{\mathcal A}}(A_1, A_2)\).

(ii)
Given a fixed degree of coherence \({{\mathcal C}_{\mathcal A}}(A_1,A_2)\) and assuming equal but arbitrary informativity \({\text {inf}}(A_1)={\text {inf}}(A_2)\), \({\Delta }(R_1)\) covaries with \(\sigma (A_1,A_2)\).
Proof
Given the assumptions of observation 1, it is shown by Schubert (2011) that the posterior probability of reliability for witness \(1\), given two witness reports \(\langle A_1^i\rangle \) and \(\langle A_2^i\rangle \) can be represented as follows:^{Footnote 25}
Hence, the following alternative representation for the benefit function for witness \(1\) given two reports \(\langle A_i^1\rangle \) and \(\langle A_i^2\rangle \) immediately follows:
This can be seen as follows:

(i)
Now given the additional assumption in (i) that \(\mathrm{inf}(A_1)=\mathrm{inf}(A_2)\), it immediately follows that there is an \(r\in (0,1)\) such that \(\Pr (A_1)=\Pr (A_2)=r\). Furthermore, given a fixed reliability parameter \(\rho \), \({\Delta }(R_1)\) is an increasing function of \(\Pr (A_1, A_2)\). Given these restrictions, \({{\mathcal C}_{\mathcal A}}(A_1,A_2)=\frac{1}{r}\cdot \Pr (A_1,A_2)\). Hence, the larger the agreement between the testimonies, the higher the impact on \({\Delta }(R_1)\), and vice versa.

(ii)
Given the additional assumption, we conclude that \({{\mathcal C}_{\mathcal A}}(A_1,A_2)=r\) for some \(r\in [0,1]\). It follows that \(\Pr (A_1,A_2)=r\cdot \Pr (A_1)\). Thus, given that we assume the coherence value to be fixed, \(\Pr (A_1,A_2)\) is an increasing function of \(\Pr (A_1)\). Thus, \({\Delta }(R_1)\) reduces to
$$\begin{aligned} {\Delta }(R_1)=\frac{\left( \frac{r}{\Pr (A_1)}1\right) \cdot (\rho ^2\rho ^3)}{\left( \frac{r}{\Pr (A_1)}1\right) \cdot \rho ^2+1} \end{aligned}$$\({\Delta }(R_1)\) is thus a decreasing function of \(\Pr (A_1)\) as is \(\sigma (A_1,A_2)=2\log _2\Pr (A_1)\). Accordingly, the larger the informations’ specificity \(\sigma (A_1,A_2)\), the higher the impact on \({\Delta }(R_1)\) and vice versa.
\(\square \)
3. Proof of Theorem 3
Theorem 3: In a scenario with two testimonies that satisfies conditions (5)(8) and fixed equal individual reliability, the following comparative assertions are true:

(i)
Given fixed equal informativity \(\mathrm{inf}(A_1)=\mathrm{inf}(A_2)\), \({\Delta }(R_1)\) covaries with \({{\mathcal C}_{\mathcal A}}(A_1,A_2)\).

(ii)
Given a fixed degree of coherence \({{\mathcal C}_{\mathcal A}}(A_1,A_2)\) and assuming equal but arbitrary informativity \(\mathrm{inf}(A_1)=\mathrm{inf}(A_2)\), \({\Delta }(R_1)\) covaries with \(\sigma (A_1,A_2)\).
Proof
I utilize the following abbreviations: \(\alpha _i=\Pr (A_i)\), \(\alpha _{1,2}=\Pr (A_1,A_2)\). Let \(\mathbf A_i, R_i, \langle A_i\rangle \) be random variables taking \(A_i\) and \(\lnot A_i\), \(R_i\) and \(\lnot R_i\), and \(\langle A_i\rangle \), \(\lnot \langle A_1\rangle \) as its values for \(i=1,2\). Then:
Utilizing the fact that \(\Pr (\langle A_i\rangle \,\, R_i, \lnot A_i)=0\), we get the following representation of \(\Pr (R_1\,\,\langle A_1 \rangle ,\langle A_2\rangle )\):
where \(\lambda :=\Pr (R_1,\langle A_1 \rangle ,\langle A_2\rangle )\) and \(\mu =\Pr (\lnot R_1,\langle A_1 \rangle ,\langle A_2\rangle )\) can be written as follows.
Putting things together we get:
Denote “\(\frac{\alpha _{1,2}}{\alpha _1\alpha _2}\)” by “\(\phi \)” and “\(\frac{\alpha '_{1,2}}{\alpha '_1,\alpha '_2}\)” by “\(\phi '\)”, then the following equivalence establishes the desired result.
\(\Pr (R_1\,\,\langle A_1 \rangle ,\langle A_2\rangle )\rho >\Pr (R_1\,\,\langle A'_1 \rangle ,\langle A'_2\rangle )\rho \)
Hence, \({\Delta }(R_1)\) is an increasing function of \(\frac{\phi _{1,2}}{\phi _1\phi _2}=\frac{\Pr (A_1,A_2)}{\Pr (A_1)\cdot \Pr (A_2)}\). Accordingly, an argument analogous to B yields the desired result. \(\square \)
4. Proof of Theorem 4
Theorem 4: In a scenario with \(n\) equivalent testimonies that satisfies conditions (5)(7) and (9), and assuming equal individual reliability, the following assertions are true irrespective of the chosen randomization parameter \(0<a<1\):

(i)
Given a fixed number \(n\) of testimonies, \({\Delta }(R_1)\) is strictly increasing in inf\((A_i)\).

(ii)
Given a fixed degree of informativity inf\((A_i)\), \({\Delta }(R_1)\) is a strictly increasing function of \(n\).
Proof
The following representation is shown to hold by Bovens and Hartmann (2003, p. 62) for the likelihood ratio \(x=\frac{\Pr (\langle A^i\rangle \,\,\lnot A)}{\Pr (\langle A^i\rangle \,\,A)}=\frac{a\overline{\rho }}{\rho +a\overline{\rho }}\), where \(\rho =\Pr (R_i)\) and \(\alpha _i=\Pr (A_i)\):

(i)
Assume that the number of testimonies is fixed. In order to show that this is a strictly decreasing function of \(\Pr (A_i)\), we calculate the first derivate and show that it always exceeds zero. Let \(f(\alpha _i):={\alpha _i\overline{x}}\) and \(g(\alpha _i):={\alpha _i+\overline{\alpha }_i x^n}\), then
$$\begin{aligned} \frac{\partial \Pr ^*(R_1)}{\partial \alpha _i}=\frac{f'(\alpha _i)g(\alpha _i)f(\alpha _i)g'(\alpha _i)}{g(h)^2} \end{aligned}$$(16)where \(f'(\alpha _i)=1x\) and \(g'(\alpha _i)=1x^n\). Accordingly, inserting these into the above equation yields the following result.
$$\begin{aligned} \frac{\partial \Pr ^*(R_1)}{\partial \alpha _i} \ = \ \frac{x^nx^{n+1}}{({\alpha _i+\overline{\alpha }_i x^n})^2}>0 \end{aligned}$$(17)With \(x\in \left( 0,1\right) \), this is positive irrespective of the chosen randomization parameter \(a\). Hence, given maximal coherence, the posterior probability of reliability again is strictly decreasing in the propositions priors.

(ii)
Now assume that the prior probability \(\alpha _i\) is fixed. It can easily be seen that if \(n\) increases, the denominator of \(\Pr ^{*}(R_1)\) decreases since \(x<1\). Accordingly, \(\Pr ^{*}(\alpha _i)\) must be an increasing function of \(n\) for fixed informativity.
\(\square \)
Rights and permissions
About this article
Cite this article
Schippers, M. Coherence, striking agreement, and reliability. Synthese 191, 3661–3684 (2014). https://doi.org/10.1007/s1122901404880
Received:
Accepted:
Published:
Issue Date:
Keywords
 Coherence
 Probability
 Reliability
 [Striking] agreement
 Bayesian epistemology