Conditionalization: Rational Belief Revision

What is Conditionalization?

At its heart, conditionalization is a simple and intuitive idea. When you learn a new piece of evidence $E$ with absolute certainty, your belief system should update in two steps:

1. All possibilities that are inconsistent with $E$ are eliminated. You assign them a credence of zero.
2. The credences for all remaining possibilities (those consistent with $E$) are proportionally rescaled so that they once again sum to 1, preserving their relative probabilities.

This process ensures that no probability is "wasted" on what is now known to be impossible, and the new distribution of belief reflects the new state of knowledge while respecting your prior judgments about the relative likelihood of the remaining possibilities.

This updating process is often summarized by the slogan: "Today's posterior is tomorrow's prior." The beliefs you arrive at after learning evidence today become the starting point for any future belief revisions.

From Principle to Calculation: Bayes' Theorem

The Principle of Conditionalization states that upon learning $E$, your new credence in any hypothesis $H$, let's call it ${\text{Cr}}_{new}(H)$, should become your old credence in $H$ conditional on $E$, or ${\text{Cr}}_{old}(H\mid E)$.

Formally, this is written as:

This is the normative rule. But how do we calculate the value on the right-hand side? The definition of conditional probability gives us the Ratio Formula:

And for more complex cases, the tool for this calculation is Bayes' Theorem. It's crucial to distinguish the principle (Conditionalization) from the tool (Bayes' Theorem). The principle tells you what your new belief state should be; the theorem tells you how to compute it from your prior beliefs.

Given your background as a statistician, you are already familiar with the mechanics of Bayes' Theorem. The following video provides a superb intuitive and visual explanation that reinforces why this formula is the natural way to perform the calculation dictated by the Principle of Conditionalization.

The Justification: The Diachronic Dutch Book

We've seen what conditionalization is, but why should a rational agent follow this specific rule for updating beliefs? Just as the standard (synchronic) Dutch Book argument provides a justification for adhering to the probability axioms, a diachronic Dutch Book argument (or "Dutch Strategy") provides a justification for Conditionalization.

The argument shows that if you commit to any updating rule other than Conditionalization, a clever bookie can offer you a series of bets over time that you will deem fair at each step, yet which will lead you to a guaranteed loss. Violating Conditionalization is thus a form of diachronic incoherence—your beliefs at different points in time do not fit together rationally.

This argument provides a powerful, pragmatic justification for Conditionalization, framing it not just as an arbitrary rule, but as a necessary component of dynamic rational coherence.

Beyond Strict Conditionalization: Uncertain Evidence

The Principle of Conditionalization as we've discussed it—often called "strict conditionalization"—assumes that you learn your evidence $E$ with absolute certainty ($\text{Cr}(E)$ becomes 1). But what if evidence is uncertain? Often, an experience doesn't make us certain of something, but merely shifts our confidence. For example, a blurry shape in the distance might increase your belief that a friend is approaching, without making you certain.

To handle these more realistic scenarios, the philosopher Richard Jeffrey proposed a generalization known as Jeffrey Conditionalization. In this model, an experience can directly shift the probability of a proposition (or a partition of propositions) to some value less than 1, and the rest of the belief system updates to cohere with this shift while preserving relevant conditional probabilities.

The existence of Jeffrey Conditionalization and ongoing debates around other challenges (like the problem of old evidence) show that while strict conditionalization is the foundational rule, the project of building a complete epistemology of belief revision is a rich and active area of philosophical inquiry.

Conclusion

Today we moved from a static to a dynamic picture of rational belief. We have seen how Bayesianism provides a normative rule for how our beliefs ought to change in the face of new evidence.

Key Takeaways:

• Conditionalization is the normative rule for rational belief revision. It states that upon learning evidence $E$, your new credence in a hypothesis $H$ should become your prior conditional credence, $\text{Cr}(H\mid E)$.
• Bayes' Theorem is the primary computational tool used to calculate the posterior probabilities required by Conditionalization.
• The justification for this rule comes from the Diachronic Dutch Book argument, which shows that any other updating plan leads to a form of incoherence over time, leaving you vulnerable to guaranteed losses.
• Strict Conditionalization is an idealized model. Generalizations like Jeffrey Conditionalization have been developed to handle more realistic cases of uncertain evidence.

We have now established both synchronic (Probabilism) and diachronic (Conditionalization) norms for rational belief. However, the entire updating process is sensitive to the starting point. Where do our initial beliefs—our priors—come from? Are all starting points equally rational as long as they are coherent? This brings us to one of the most significant and debated topics in Bayesian philosophy, which we will tackle in our next lesson: the problem of the priors.