Bayesian Convergence and Scientific Objectivity

The Subjectivist's Gambit: Washing Out the Priors

The charge against subjectivism is that it leads to a kind of epistemic relativism. If you and I start with vastly different priors about a scientific hypothesis, we might interpret the same evidence in completely different ways. You might see a piece of data as confirmation, while I see it as disconfirmation. If both our views are "rational," what hope is there for consensus?

The convergence-of-opinion theorem (also known as the "merging-of-opinions" or "washing out of the priors" theorem) offers a formal mathematical escape route. The core idea is that, under certain conditions, as two agents with different initial beliefs observe the same sequence of evidence, their posterior probabilities for a given hypothesis will inevitably converge. In the long run, the weight of accumulated evidence is supposed to swamp the influence of their initial subjective priors.

Let's start with a concise academic framing of this idea.

Now, let's turn to a more detailed explanation of how and why this happens.

The key insight is that evidence has more confirmatory power when it is surprising. An agent who starts with a low prior for a hypothesis $H$ will also, logically, have a low prior for evidence $E$ that is predicted by $H$. When that evidence $E$ is observed, their belief in $H$ must increase dramatically to maintain coherence. In contrast, the agent who already had a high prior for $H$ was already expecting $E$, so observing it provides only a minor boost to their belief. Over time, this differential updating pulls divergent beliefs toward each other.

The following image visualizes this process. Different agents start with different prior distributions (dashed lines), but after observing the same data, their posterior distributions (solid lines) become much more similar.

The Mechanism: The Overwhelming Power of the Bayes Factor

To understand this more formally, we can look at the "odds form" of Bayes' rule:

In words:

The Prior Odds capture the initial subjective disagreement. The Bayes Factor measures the strength of the evidence. It's the ratio of how probable the evidence is if the hypothesis is true, versus how probable it is if the hypothesis is false.

As a stream of evidence $E_1,E_2,...,E_n$ comes in, we multiply by the Bayes Factor for each piece of evidence. If the evidence consistently supports $H$, the cumulative Bayes Factor will become an enormous number. If it consistently undermines $H$, it will approach zero. In either case, after enough evidence, the cumulative Bayes Factor will be so extreme that it will dominate the initial Prior Odds, forcing the Posterior Odds of all observers toward the same value.

This next image illustrates the relationship. When the evidence is weak (Likelihood Ratio L is close to 1), the posterior is sensitive to the prior. But when evidence is strong (L is very high or low), the posterior is forced to 1 or 0, almost regardless of the prior.

Evaluating the Promise: Is Convergence Enough for Objectivity?

The convergence theorem is mathematically elegant and philosophically reassuring. But does it work in the real world of science? When we "evaluate" the theorem, we must look at its assumptions and limitations.

Let's formalize these powerful criticisms:

1. The "In the Long Run, We're All Dead" Problem: Convergence is an asymptotic property. It guarantees agreement eventually, but "eventually" could be after a quantity of evidence that is impossible to gather in a human lifetime, or even within the lifetime of a scientific paradigm. Scientific decision-making happens in the short and medium run, where initial priors can still hold significant sway.

2. Disagreement about Likelihoods: The theorem assumes that while agents can disagree on priors $P(H)$, they agree on the likelihoods $P(E|H)$. This assumes they agree on all the auxiliary hypotheses and background theories that connect the main hypothesis $H$ to the evidence $E$. In many real scientific controversies (e.g., interpreting quantum mechanics, assessing climate models), the debate is precisely about these connecting assumptions, meaning the likelihoods themselves are a point of contention. If agents don't agree on the Bayes Factor, there is no guarantee their beliefs will converge.

3. The Problem of Dogmatism: Convergence only works if agents are open-minded. If an agent assigns a prior probability of 0 to a hypothesis, no amount of evidence can ever resurrect it (since $P(H|E)=0$ if $P(H)=0$). A dogmatic belief, once set to zero, is forever immune to evidence. This is why many Bayesians insist on the principle of Regularity (or Cromwell's Rule), which forbids assigning priors of 0 or 1 to any contingent proposition.

The Stanford Encyclopedia of Philosophy entry on Bayesian Epistemology provides a good summary of these issues.

Conclusion

So, where does this leave us? Does the convergence-of-opinion theorem successfully rescue scientific objectivity for the subjective Bayesian? The answer, like most interesting philosophical answers, is nuanced.

Key Takeaways:

• The convergence-of-opinion theorem is the subjective Bayesian's primary argument for objectivity. It mathematically demonstrates how agents with different priors will have their beliefs converge as they observe the same evidence.
• The mechanism is the overwhelming power of the Bayes Factor. Over time, the cumulative strength of evidence is expected to swamp the influence of initial prior odds.
• However, the theorem's power as a practical guarantee of scientific objectivity is limited by several critical assumptions:
  • It guarantees convergence only in the long run, which may be too long to be relevant for actual scientific practice.
  • It assumes agents agree on the likelihoods ($P(E|H)$), which is often not the case in complex scientific debates.
  • It does not work for dogmatic agents who assign priors of 0 or 1.

The theorem doesn't prove that objectivity is achieved in practice, but it provides a powerful, formal model of how intersubjective agreement can emerge from subjective starting points. It shows that the Bayesian framework has the internal resources to explain consensus-building, even if it cannot command it on any given Tuesday.

This lesson concludes our exploration of the core tenets and debates within Bayesian epistemology. We've moved from the foundations of probabilistic belief to the dynamics of belief updating and the deep problem of where those beliefs originate.

Now, we will pivot. So far, we have treated evidence as something we simply observe. But what is the difference between seeing and doing? In the next module, we will begin our journey into causal analysis, starting with the philosophical shift required to move from mere statistical correlation to genuine causal claims. Our first topic will be how interventionist theories of causation sought to overcome the limitations of older views.