---
title: Spohn's Ranking Functions (OCFs)
url: https://www.emergentmind.com/topics/spohn-s-ranking-functions
type: topic
---

# Spohn's Ranking Functions (OCFs)

Spohn’s ranking functions, also called **ordinal conditional functions** (OCFs), are epistemic-state representations that assign ordinal degrees of disbelief, implausibility, or surprise to possible worlds and extend those values to propositions by minimization over their supporting worlds. In the finite settings emphasized in formal, comparative, and computational work, they serve as a calculus for plausibility ordering, deductively closed belief, conditional acceptance, conditioning, iterated revision, and forgetting. They also admit explicit translations into possibility measures and principled, though non-invertible, transformations to and from probability measures [1304.1118], [1705.07226], [2508.21441], [1301.6699].

## 1. Formal structure and semantic interpretation

In finite presentations, a ranking function assigns nonnegative integers to worlds, sometimes with an added value \(\infty\) for impossibility. RankPL defines a ranking function as
\[
\kappa:\Omega\to \mathbb{N}\cup\{\infty\},
\]
with
\[
\kappa(\emptyset)=\infty,\qquad \kappa(A)=\min\{\kappa(w)\mid w\in A\}
\]
for nonempty events \(A\), and the normalization condition
\[
\kappa(\Omega)=0.
\]
The 2025 forgetting framework presents OCFs as
\[
\kappa:\Omega\to\mathbb{N}
\]
with \(\kappa^{-1}(0)\neq\emptyset\), which is the same normalization requirement in equivalent form: at least one world is maximally plausible [1705.07226], [2508.21441].

Dubois and Prade use a finite Boolean algebra \(B\) induced by a partition \(\{A_1,\dots,A_m\}\) of \(\Omega\), and impose three structural conditions: constancy of \(\kappa\) on each cell \(A_i\), existence of at least one cell with rank \(0\), and event evaluation by
\[
\kappa(A)=\min\{\kappa(\omega)\mid \omega\in A\}.
\]
In that formulation, \(\kappa(\omega)\) is interpreted as a **degree of impossibility** of world \(\omega\), and \(\kappa(A)\) gives the least degree of impossibility among the worlds in \(A\). A proposition is therefore evaluated by its most plausible realization, but on a disbelief scale rather than a plausibility scale. The paper states the semantic consequence explicitly: “\(\kappa(A)=0\) means \(A\) is completely possible” [1304.1118].

This min-based semantics induces the characteristic algebra of rankings. For disjoint events \(A\) and \(B\),
\[
\kappa(A\cup B)=\min(\kappa(A),\kappa(B)).
\]
The same property is used in the forgetting paper in the propositional form
\[
\kappa(A\vee B)=\min\{\kappa(A),\kappa(B)\}.
\]
Unlike probability, a disjunction is as unsurprising as its least surprising disjunct. RankPL therefore describes rankings as a qualitative calculus of uncertainty in which \(0,1,2,\dots\) express increasing surprise, while \(\infty\) marks impossibility [1705.07226], [2508.21441].

## 2. Belief, conditionals, and epistemic equivalence

A ranking function induces a propositional belief set from its rank-\(0\) worlds. The forgetting framework defines
\[
Bel(\kappa)=Th(\{\omega\mid \kappa(\omega)=0\}),
\]
and equivalently
\[
Bel(\kappa)=Cn\!\left(\bigvee_{\omega:\kappa(\omega)=0}\omega\right).
\]
Its key belief criterion is
\[
A\in Bel(\kappa)\quad\text{iff}\quad \kappa(\overline A)>0.
\]
Thus a proposition is believed exactly when every most plausible world satisfies it [2508.21441].

RankPL presents the same general idea with an explicit firmness parameter: \(A\) is believed with firmness \(x>0\) iff
\[
\kappa(\overline A)>x.
\]
Under this criterion, belief is deductively closed under conjunction: if \(A\) and \(B\) are each believed with firmness \(x\), then so is \(A\cap B\). Both RankPL and Giang–Shenoy emphasize this as a distinguishing feature relative to probabilistic belief, and both connect it to the Lottery Paradox [1705.07226], [1301.6699].

Ranking functions also support conditional acceptance. The forgetting framework treats \((B\mid A)\) as accepted exactly when the best verifying worlds are more plausible than the best falsifying worlds:
\[
\kappa\models (B\mid A)\quad\text{iff}\quad \kappa(AB)<\kappa(A\overline B).
\]
Ordinary propositional belief is recovered as the special case
\[
\kappa\models A\quad\text{iff}\quad \kappa\models (A\mid \top).
\]
This is one reason OCFs are treated as full epistemic states rather than as flat belief sets: they validate propositional beliefs and conditionals within the same plausibility ordering [2508.21441].

The literature summarized here also distinguishes several notions of sameness between rankings. Two OCFs are epistemically equivalent if they preserve all pairwise world orderings:
\[
\kappa\cong\kappa' \quad\text{iff}\quad
\forall \omega_1,\omega_2\in\Omega:\ 
\kappa(\omega_1)\le \kappa(\omega_2)\ \text{iff}\ 
\kappa'(\omega_1)\le \kappa'(\omega_2).
\]
A stronger arithmetic relation is **linear equivalence**:
\[
\kappa_1\equiv_x \kappa_2
\quad\text{iff}\quad
\exists q\in\mathbb{Q}_{>0}\text{ such that }\kappa_2=q\cdot\kappa_1.
\]
If \(\kappa_2=q\cdot\kappa_1\), then for any formula \(A\),
\[
\kappa_2(A)=q\cdot\kappa_1(A).
\]
This scaling property becomes important in forgetting, where some operators are evaluated against a ranking-specific linear-equivalence postulate \((LE^\times)\) [2508.21441].

Giang and Shenoy also define a derived **Spohnian belief function**
\[
\beta(A)=
\begin{cases}
-\delta(A) & \text{if } \delta(A)>0,\\
\delta(-A) & \text{otherwise.}
\end{cases}
\]
Their purpose is again to connect ordinal disbelief with plain belief in a deductively closed form [1301.6699].

## 3. Conditioning, uncertain evidence, and iterated revision

Ordinary conditioning is the basic update mechanism. In RankPL, conditional rank is defined by
\[
\kappa(A\mid B)=
\begin{cases}
\kappa(A\cap B)-\kappa(B) & \text{if }\kappa(B)\neq\infty,\\
\infty & \text{otherwise,}
\end{cases}
\]
and the conditioned ranking is \(\kappa_B\) with
\[
\kappa_B(A)=\kappa(A\mid B).
\]
At the world level, Dubois and Prade write, for \(\omega\in A\),
\[
\kappa(\omega\mid A)=\kappa(\omega)-\kappa(A).
\]
In both formulations, conditioning renormalizes the accepted event so that its best worlds receive rank \(0\); in RankPL the complement is shifted to \(\infty\), so observation is a very strong revision operation [1705.07226], [1304.1118].

Spohn’s treatment of uncertain evidence is different from ordinary conditioning. Dubois and Prade present the \((A,n)\)-conditionalization
\[
\kappa(\omega\mid (A,n))=
\begin{cases}
\kappa(\omega\mid A), & \text{if }\omega\in A,\\[4pt]
n+\kappa(\omega\mid \overline A), & \text{if }\omega\in \overline A.
\end{cases}
\]
Its effect is to penalize \(\overline A\) by an additional amount \(n\). The larger \(n\), the stronger the support for \(A\). In their comparison with possibilistic Jeffrey-style updating, this is the distinguishing mechanism: uncertain evidence is represented by rank shifts, not by averaging posterior states [1304.1118].

RankPL discusses two generalized ranking-theoretic revision schemes for noisy and iterated evidence. **J-conditioning** revises by finite firmness:
\[
\kappa_{A\rightarrow x}(B)=\min\bigl(\kappa(B\mid A),\ \kappa(B\mid \overline A)+x\bigr).
\]
Its effect is that \(A\) becomes believed with firmness \(x\), not with infinite firmness. Because \(\overline A\) is moved upward only finitely, later evidence can reverse the update. **L-conditioning** measures impact rather than target firmness:
\[
\kappa_{A\uparrow x}(B)=\min\bigl(\kappa(A\cap B)-y,\ \kappa(\neg A\cap B)+x-y\bigr),
\qquad
y=\min(\kappa(A),x).
\]
RankPL highlights two properties of L-conditioning:
\[
(\kappa_{A\uparrow x})_{\overline A\uparrow x}=\kappa,
\qquad
(\kappa_{A\uparrow x})_{B\uparrow x}=(\kappa_{B\uparrow x})_{A\uparrow x}.
\]
These are its reversibility and commutativity properties, respectively [1705.07226].

Dubois and Prade also give a general partition-based extension of Spohn updating:
\[
\pi(\omega\mid \{(A_j,\alpha_j)\}_{j=1}^n)=\alpha_i\,\pi_1(\omega\mid A_i)
\quad \text{for } \omega\in A_i.
\]
In the singleton-partition case with \(\alpha_i=\pi_2(\omega)\), this yields
\[
\pi(\omega\mid \pi_2)=\pi_2(\omega),\qquad \forall \omega\in\Omega.
\]
In that extreme case, generalized Spohn updating simply substitutes the new possibility distribution for the old one [1304.1118].

## 4. Possibility-theoretic and probabilistic correspondences

A central comparative result is the explicit connection between OCFs and possibility theory. Dubois and Prade define
\[
N_\kappa(A)=1-e^{-\kappa(A)},
\qquad
\Pi_\kappa(A)=e^{-\kappa(A)},
\qquad
\pi_\kappa(\omega)=e^{-\kappa(\omega)}.
\]
They identify \(N_\kappa\) as a necessity measure and \(\pi_\kappa\) as the associated possibility distribution. At the set level,
\[
\Pi(A)=\sup_{\omega\in A}\pi(\omega).
\]
In the finite setting they consider, this gives an order-reversing exponential translation from disbelief to possibility: lower rank means higher possibility. Because \(\kappa(\omega)\in\mathbb N\), all \(\pi_\kappa(\omega)\) are strictly positive, so in that presentation “nothing is considered as fully impossible” [1304.1118].

Under this translation, certain conditioning coincides exactly between the two formalisms. Dubois and Prade write
\[
\pi_\kappa(\omega\mid A)=\frac{\pi_\kappa(\omega)}{\Pi_\kappa(A)}
\quad \text{if } \omega\in A,
\]
which matches the possibilistic conditioning rule
\[
\pi(\omega\mid B)=
\begin{cases}
\dfrac{\pi_1(\omega)}{\Pi_1(B)}, & \text{if }\omega\in B,\\[6pt]
0, & \text{otherwise.}
\end{cases}
\]
At the set level,
\[
\Pi(A\mid B)=\frac{\Pi_1(A\cap B)}{\Pi_1(B)}.
\]
The relation becomes more delicate for uncertain information. Dubois and Prade’s own Jeffrey-like possibilistic update is
\[
[\pi_1\mid \pi_2](\omega)=
\min\!\left(
\pi_2(\omega),\
\frac{\pi_1(\omega)}{\Pi_1(B_{2\,\pi_2(\omega)})}
\right),
\]
with
\[
B_{2\,\pi_2(\omega)}=\{\omega'\mid \pi_2(\omega')\ge \pi_2(\omega)\}.
\]
Their main conclusion is that Spohn’s rule is not simply another Jeffrey rule. If \(\pi_2\le \pi_1\), both rules produce \(\pi_2\); if \(\pi_2\ge \pi_1\), the possibilistic rule yields \([\pi_1\mid \pi_2]=\pi_1\), whereas Spohn’s rule may still replace the old state by the new one. Dubois and Prade interpret this as a difference between productive refinement and priority to the new information [1304.1118].

Giang and Shenoy study the relation to probability through explicit transformations between probability distributions and Spohnian disbelief functions. Their forward transformation \(T:\mathcal P\to\Delta\) is constrained by **ordinal congruence I**:
\[
p(A)\ge p(B)\implies T(p)(A)\le T(p)(B).
\]
To maximize retained ordinal information, they introduce **leap indices** for a non-increasing probability sequence \(q_1\ge\cdots\ge q_n\):
\[
L_Q=\{\,i\mid q_i>\sum_{j=i+1}^n q_j\,\}.
\]
If worlds are ordered so that \(p(w_1)\ge\cdots\ge p(w_n)\), the transformation uses a disbelief counter \(r\) and remaining mass \(M\): assign the current level \(r\), subtract the current probability from \(M\), and increment \(r\) whenever \(p_i>M\). Their Theorem 1 states that this \(T\) is a **least-coarse congruent** probability-to-disbelief transformation [1301.6699].

The reverse transformation \(S:\Delta\to\mathcal P\) is governed by **ordinal congruence II**:
\[
\delta(A)<\delta(B)\implies S(\delta)(A)>S(\delta)(B).
\]
If \(k_i\) is the number of worlds in disbelief stratum \(i\), then for \(\delta(w)=i\),
\[
S(\delta)(w)=\frac{1}{(k_0+1)(k_1+1)\cdots(k_i+1)}\cdot Z,
\]
where \(Z\) is the normalization constant. This yields equal probability within each disbelief stratum and multiplicative discounting across strata. Giang and Shenoy prove that \(S\) is congruent, but also stress that the reverse direction is not unique and that
\[
S(T(p))\neq p
\quad\text{in general.}
\]
They also establish a dynamic compatibility result:
\[
T(S(\delta)(.\mid A))=D(\delta(. \mid A)),
\]
where \(D\) is densification. Thus transformation and conditioning commute up to removal of empty disbelief levels [1301.6699].

## 5. Programming-language realization and computational use

RankPL makes ranking theory operational by using ranking functions as the semantic domain of a qualitative probabilistic programming language. Its central semantic object is a transformation
\[
D\llbracket s\rrbracket
\]
from prior rankings over program states to posterior rankings over program states. Program states are valuations \(\sigma\), the set of proper rankings is \(K\), and there is a special failure ranking \(\kappa_\infty\) assigning \(\infty\) to every valuation; the semantic codomain is therefore
\[
K^*=K\cup\{\kappa_\infty\}.
\]
The language extends a small imperative core with three ranking-specific constructs: ranked choice, observation, and rank expressions [1705.07226].

The ranked choice
\[
\{s_1\}\langle e\rangle\{s_2\}
\]
treats \(s_1\) as the normal branch and \(s_2\) as a surprising branch whose cost is the value of \(e\). Its denotation is
\[
D\llbracket \{s_1\}\langle e\rangle\{s_2\}\rrbracket(\kappa)=||\lambda||,
\]
where
\[
\lambda(\sigma)=
\min\bigl(
D\llbracket s_1\rrbracket(\kappa)(\sigma),\
D\llbracket s_2\rrbracket(\kappa)(\sigma)+\sigma_\kappa(e)
\bigr).
\]
Observation implements conditioning:
\[
D\llbracket observe\ b\rrbracket(\kappa)=
\begin{cases}
\kappa_\infty & \text{if }\kappa=\kappa_\infty \text{ or }\kappa(b)=\infty,\\[4pt]
\kappa_b & \text{otherwise.}
\end{cases}
\]
Conditionals and loops are defined denotationally by splitting, conditioning, recombining, and iterating rankings rather than by numerical probability propagation [1705.07226].

The paper’s examples illustrate how this semantics is used. A toy program with nested ranked choices yields three outcomes, \(x=10\) of rank \(0\), \(x=20\) of rank \(1\), and \(x=30\) of rank \(2\); adding an observation \(y>1\) removes the first outcome and shifts the others down to ranks \(0\) and \(1\). A full-adder diagnosis example uses surprise penalties on component failures and an observation of anomalous input-output behavior to rank explanations abductively; for the specified observation, the posterior ranking has a unique rank-\(0\) explanation in which one particular XOR gate fails. A robot-localization example uses L-conditioning with strength \(1\) to incorporate noisy sensor readings without rendering inconsistent alternatives unrevisable; after misleading observations, later evidence restores the actual location to rank \(0\) [1705.07226].

RankPL also emphasizes implementation. Its interpreter is described as faithful to the denotational semantics and uses a **most-plausible-first execution strategy**, exploring alternatives in ascending order of rank. This exploits a practical property of ranking semantics: in diagnosis and abduction, the most plausible outcomes are typically the primary target [1705.07226].

## 6. Forgetting, contraction, and epistemic-state change

The 2025 framework treats forgetting as an operation on epistemic states with richer structure than a belief set. In that setting, forgetting may affect propositional beliefs, accepted conditionals, relevance structure, and the plausibility ordering itself. Five abstract kinds of epistemic forgetting are distinguished for a contingent formula \(A\): **contraction**, **ignoration**, **revocation**, **marginalization**, and **conditionalization**. The OCF instantiation then yields seven concrete forgetting operators [2508.21441].

| Operator | Defining pattern | Principal role |
|---|---|---|
| OCF-marginalization | \(\kappa^\circ_A=\kappa_{\Sigma\setminus \Sig{A}}\) | variable forgetting / language reduction |
| Lifted marginalization | \(\kappa^\circ_A=(\kappa_{\Sigma\setminus \Sig{A}})_{\uparrow\Sigma}\) | variable forgetting while preserving the original signature |
| Conditionalization | \(\kappa^\circ_A=\kappa\mid \overline A\) | revocative forgetting by moving to the \(\neg A\) context |
| c-Ignoration | c-contraction with \(\gamma=\kappa(\overline A)-\kappa(A)\) | suspension of judgment between \(A\) and \(\neg A\) |
| c-Revocation | c-contraction with \(\gamma>\kappa(\overline A)-\kappa(A)\) | forgetting \(A\) by accepting \(\neg A\) |
| Minimal c-contraction | c-contraction with \(\gamma=\kappa(\overline A)\) | AGM-style contraction |
| Non-minimal c-contractions | c-contractions with \(\gamma\neq \kappa(\overline A)\) | broader contraction family beyond the minimal case |

OCF-marginalization to a subsignature \(\Sigma'\subseteq\Sigma\) is defined by
\[
\kappa_{\Sigma'}(\omega')=
\min\{\kappa(\omega)\mid \omega\in\Omega_\Sigma\text{ and }\omega\models \omega'\}.
\]
Its associated forgetting operation is
\[
\kappa^\circ_A=\kappa_{\Sigma\setminus \Sig{A}}.
\]
This is the semantic counterpart of variable elimination. The paper proves the exact compatibility result
\[
Bel(\kappa_{\Sigma'})=Bel(\kappa)_{\Sigma'},
\]
and therefore
\[
Bel(\kappa^\circ_A)=Bel(\kappa)\cap \mathcal L_{\Sigma\setminus \Sig{A}}.
\]
Lifted marginalization keeps the original signature but makes forgotten atoms irrelevant to plausibility by defining
\[
(\kappa)_{\uparrow\Sigma}(\omega)=\kappa(\omega^{\Sigma'}).
\]
Conditionalization, by contrast, is
\[
\kappa\mid A(\omega)=\kappa(\omega)-\kappa(A),
\qquad
\kappa^\circ_A=\kappa\mid \overline A,
\]
so it behaves more like revision by \(\neg A\) than like variable forgetting [2508.21441].

The contraction-style family is numerically defined by world-rank shifts. For contingent \(A\), a c-contraction \(\kappa-A\) is characterized by the existence of an integer \(\gamma\) such that
\[
\gamma\ge \kappa(\overline A)-\kappa(A)
\]
and
\[
(\kappa-A)(\omega)=
-\kappa(\overline A)+\kappa(\omega)+
\begin{cases}
\gamma & \text{if }\omega\models A,\\
0 & \text{if }\omega\models \overline A.
\end{cases}
\]
Different choices of \(\gamma\) yield ignoration, revocation, minimal contraction, and non-minimal contraction. Minimal c-contraction is singled out as the canonical AGM-style forgetting operator; c-ignoration is the unique strategy that makes at least one \(A\)-world and one \(\neg A\)-world rank \(0\), thereby believing neither \(A\) nor \(\neg A\) [2508.21441].

The postulate analysis provides a sharp classification. OCF-marginalization satisfies AGM(1), AGM(3), AGM(5), AGM(6), \((W)\), \((wC)\), \((sC)\), \((CP)\), \((wE)\), \((E)\), \((BE)\), \((EBE)\), \((OI)\), \((PP)\), \((NP)\), \((EP)\), \((BP)\), and \((LE^\times)\), but violates AGM(2), AGM(4), AGM(7), and \((CF)\). Minimal c-contraction satisfies AGM(1)–AGM(7), \((CF)\), \((BE)\), \((EBE)\), and \((LE^\times)\), but violates the persistence and weakening-style postulates. Conditionalization and c-revocation are close to revision by negation; c-ignoration models neutrality about \(A\); non-minimal c-contractions lack the rationality guarantees of the minimal variant. The framework’s own synthesis is therefore that **OCF-marginalization** is the canonical realization of variable forgetting, whereas **minimal c-contraction** is the canonical realization of AGM contraction [2508.21441].

Across these applications, ranking functions are treated not merely as numeric annotations on worlds but as structured epistemic states. That is why the same formalism supports propositional belief, conditional acceptance, revision, contraction, marginalization, and forgetting, and why two states with the same propositional belief set can still behave differently under forgetting or update if their plausibility orderings differ [2508.21441].

Source: https://www.emergentmind.com/topics/spohn-s-ranking-functions