## Definition of Relevance

In the previous essay in this series, we introduced the basic ideas and terminology of Bayesian argumentation, including the concept of **relevance**.

To review, **a premise $B$ is relevant to conclusion $A$** (in the mind of the subject) **iff**:

$$ P(A|B) β P(A|\bar{B}) $$

If the above condition does not hold, then **π΅ is irrelevant to π΄**. It’s easy to show that this is the case if and only if π΄ and π΅ are statistically independent (proof).

## Definition of Support and Oppose

We say that premise **supports** the conclusion (in the mind of the subject) iff the subject is **more** likely to accept the premise if they accept the conclusion. That is, **π΅ supports π΄ iff**:

$$ P(A \vert B) > P(A \vert \bar{B}) $$

If the subject is **less** likely to accept π΄ if they accept π΅ ($P(A \vert B) < P(A \vert \bar{B})$), it follows that they are **more likely to accept not π΄**, in which case we say that the premise **opposes** the conclusion. That is, **π΅ opposes π΄ iff:**

$$ P(A|B) < P(A|\bar{B}) βΊ P(\bar{A}|B) > P(\bar{A}|\bar{B}) $$

## Quantifying Relevance

We have defined the term **relevant** as a binary attribute. But we often talk about degrees or relevance. This can be measured as the difference between $P(A \vert B)$ and $P(A \vert \bar{B})$.

### Definition of Relevance

**The relevance of π΅ to π΄ is**:

$$ \label{1} R(A,B) = P(A|B) - P(A|\bar{B}) \tag{1} $$

The relevance will be negative if π΅ opposes π΄. Now recall that if π΅ opposes π΄ it supports $\bar{A}$. And it’s also the case that if π΅ opposes π΄, $\bar{B}$ supports π΄! In fact, the following are all equal:

$$ R(A,B) = -R(\bar{A},B) = -R(A,\bar{B}) = R(\bar{A},\bar{B}) $$

### Acceptance of the Premise

Now a Bayesian agent doesn’t just accept or reject premises. In a Bayesian model all beliefs are probabilities.

I’ll use the term **acceptance** to mean the probability that the user accepts a claim (the premise or conclusion). $ P(A \vert B) $ is acceptance of the conclusion given that the acceptance of the premise is 100%. And $P(A \vert \bar{B})$ is the acceptance of the conclusion given acceptance of the premise is 0%. But what will acceptance of the conclusion be if acceptance of the premise was, say, 50%?

### Relevance as Slope

It turns out that the relationship between acceptance of the premise and conclusion is linear, and **relevance is the slope of the line relating acceptance of the premise with acceptance the conclusion**.

$$ \label{2} P(A) = P(A|\bar{B}) + P(B)R(A,B) \tag{2} $$

This relationship is illustrated in the chart below:

The horizontal axis is the **posterior** belief $P’(B)$, and the vertical axis is the **posterior** belief $P’(A)$. The line intersects the vertical at $P(A \vert \bar{B})$ – the subject’s belief were they to completely reject π΅. The posterior belief in π΄ increases linearly as the posterior belief in π΅ increases, to the point that the subject completely accepts π΅, and the belief in $P’(A)$ has the maximum possible value $P(A \vert B)$. The prior beliefs, $P(A)$ and $P(B)$, are a point on this line.

## Conditional Relevance and Corelevance

Relevance exists in the context of the subject’s other prior beliefs. For example, if (πΆΜ
) *the car is out of gas*, and also ($\bar{B}$) *the battery is dead*, then both of these are good reasons to believe (π΄Μ
) *the car won’t start*. Yet neither is relevant on its own by the definition of relevance given above. Given that the car is out of gas, it makes no difference whether the battery is dead or not: the car won’t start anyway. In other words, ($\bar{B}$) *the battery is dead* is irrelevant to (π΄Μ
) *the car won’t start* given (πΆΜ
) *the car is out of gas*.

But if the subject believes (πΆ) *the car has gas*, then ($\bar{B}$) the battery is dead will probably be relevant. When accepting one premise causes another premise to become relevant, we say that the premises are **corelevant**. If a premise is corelevant with some unexpressed premise, we can say that the premise is **conditionally relevant**.

### Definition of Conditional Relevance

To define corelevance mathematically, we need to first define the **conditional relevance** of π΅ to π΄ given πΆ, $R(A,B \vert C)$:

$$ R(A,B|C) = P(A|B,C) - P(A|\bar{B},C) $$

### Definition of Corelevant

Then π΅ and πΆ are corelevant to π΄ if:

$$ R(A,B|C) β R(A,B|\bar{C}) $$

### Quantifying Corelevance

We can measure the magnitude of the correlevance as the difference:

$$ CR(A;B,C) = R(A,B|C) - R(A,B|\bar{C}) $$

It’s easy show that co-relevance is symmetrical (proof).

$$ CR(A;B,C) = CR(A;C,B) $$

### Counterfactual Relevance

Unfortunately, this definition of conditional relevance still doesn’t capture the the common notion of “relevance” very well, because we can almost always find some second premise that makes the premise conditionally relevant. For example, the premise (π») *The car has a hood ornament* may not seem relevant to (π΄) *the car will start*, but it is conditionally relevant given the premise (π) *The car is powered by a magical hood ornament*.

Of course, π is pretty implausible – $P(M)$ may be infinitesimally small. But other more plausible corelevant premises may have small probabilities. For example, if the subject just filled the car with gas, they will be quite certain that (πΊ) *the car has gas* and thus $P(\bar{G})$ might be infinitesimally small. So in both cases we have corelevant premises with small prior probabilities, but a car running out of gas is something that is likely to actually happen in many similar scenarios, even if not this particular one.

Accounting for the difference in relevance in these two cases takes us into the metaphysical realm of modal logic, possible worlds, counterfactuals, and other difficult epistemological questions, that we won’t try to answer here.

## Next in this Series

In our introductory example, we claimed that *he has a pulse* is relevant to the conclusion *this is a good candidate for the job*. But it is obviously not a very good argument. Why not?

Obviously, because the subject probably already assumed that the candidate had a pulse. Relevance doesn’t say anything about the subject’s **actual** prior degree of belief in the premise or conclusion. In the next essay in this series, we will show that, because the subject already believes that *he has a pulse* is true, it is a **necessary but not sufficient** premise for the conclusion *this is a good candidate for the job*.

## Proofs

### Proof 1

**Jeffrey’s Rule of Conditioning**

If a Bayesian agent acquires information that has no direct effect other than to increase their belief in $B$ from prior $P(B)$ to posterior $P’(B)$, then their posterior belief $P’(A)$ changes according to

$$ P’(A) = P(A|\bar{B}) + P’(B)R(A,B) $$

**Proof:**

First, the above assumptions mean that a change in $P(B)$ does not change the **conditional** probabilities $P(A \vert B)$ and $P(A \vert \bar{B})$. So:

$$ \begin{aligned} P’(A|B) &= P(A|B)\cr P’(A|\bar{B}) &= P(A|\bar{B}) \end{aligned} $$

This means that the relevance doesn’t change either:

$$ R’(A,B) = P’(A|B) - P’(A|\bar{B}) = P(A|B) - P(A|\bar{B}) = R(A,B) $$

Now, the equality $\eqref{2}$ holds for any probability distribution, including the posterior probability distribution $P’$. So:

$$ P’(A) = P’(A|\bar{B}) + P’(B)R’(A,B) ~~~~~ \text{Formula }\eqref{2} $$

And since the posteriors $P’(A|\bar{B})$ and $R’(A,B)$ are the same as the priors:

$$ \begin{aligned} P’(A) &= P’(A|\bar{B}) + P’(B)R’(A,B) \cr &= P(A|\bar{B}) + P’(B)R(A,B) \end{aligned} $$

This formula can be rearranged to express Jeffrey’s rule in the more typical form:

$$ \begin{aligned} P’(A) &= P(A|\bar{B}) + P’(B)R(A,B) \cr &= P(A|\bar{B}) + P’(B)( P(A|B) - P(A|\bar{B}) ) \cr &= P(A|\bar{B}) + P’(B)P(A|B) - P’(B)P(A|\bar{B}) ) \cr &= P(A|\bar{B}) + P’(B)P(A|B) - (1 - P’(\bar{B}))P(A|\bar{B}) ) \cr &= P’(B)P(A|B) + P’(\bar{B})P(A|\bar{B}) \end{aligned} $$

### Proof 2

**(Ir)Relevance Implies (In)Dependence**

$$ R(A,B) = 0 βΊ P(A) = P(A|B) = P(A|\bar{B}) $$

**Proof:**

If $R(A,B) = 0$, then $P(A \vert B) = P(A \vert \bar{B})$, because:

$$ 0 = R(A,B) = P(A|B) - P(A|\bar{B}) \iff P(A \vert B) = P(A \vert \bar{B}) $$

And it also follows that $P(A) = P(A \vert \bar{B})$, because:

$$ \begin{aligned} P(A) &= P(A|\bar{B}) + P(B)R(A,B) ~~~~~ \text{Formula } \eqref{2} \cr &= P(A|\bar{B}) \end{aligned} $$

### Proof 3

**Symmetry of Corelevance**

$$ CR(A;B,C) = CR(A;C,B) $$

**Proof:**

$$ \begin{aligned} CR(A;B,C) &= R(A,B \vert C) - R(A,B \vert \bar{C}) \cr &= ( P(A \vert B,C) - P(A \vert \bar{B},C) ) \cr &\space\space\space\space- ( P(A \vert B,\bar{C}) - P(A \vert \bar{B},\bar{C}) ) \cr &= ( P(A \vert B,C) - P(A \vert B,\bar{C}) ) \cr &\space\space\space\space- ( P(A \vert \bar{B},C) - P(A \vert \bar{B},\bar{C}) ) \cr &= ( P(A \vert C,B) - P(A \vert \bar{C},B) \cr &\space\space\space\space- ( P(A \vert C,\bar{B}) - P(A \vert \bar{C},\bar{B}) ) \cr &= R(A,C \vert B) - R(A,C \vert \bar{B}) \cr &= CR(A;C,B) \cr \end{aligned} $$