Universal Positive Almost Sure Termination (UPAST)

• Universal Positive Almost Sure Termination (UPAST) is defined as the property that every program input results in finite expected termination time, ensuring both probability‐1 termination and bounded runtime.
• It is characterized as Π⁰₃-complete in the fully probabilistic imperative while-program model, highlighting its increased complexity over standard AST and PAST.
• Multiple proof strategies, including ranking measure functions, lexicographic supermartingales, and generating function approaches, have been developed to precisely capture UPAST in various probabilistic models.

Universal Positive Almost Sure Termination (UPAST) is the property that a probabilistic program terminates in finite expected time under the relevant universal quantification induced by its semantics. In the fully probabilistic imperative while-program model of Kaminski, Katoen, Matheja, and Olmedo, UPAST is defined by requiring finite expected termination time for every input valuation, and the corresponding decision problem is shown to be $\Pi^0_3$-complete (Kaminski et al., 2015). Across the literature, the same semantic pattern reappears under model-specific names: universality may range over all inputs, all schedulers, all initial configurations, or all full rewrite sequences, but the central distinction remains the same—UPAST strengthens almost-sure termination by demanding finite expectation as well as probability-$1$ termination.

1. Definition and semantic scope

In the while-program model used for the hardness classification, a program $P$ is evaluated from an input valuation $\eta$, with termination probability

$\Pr_{P,\eta}(\downarrow)$

and expected time until termination

$E_{P,\eta}(\downarrow).$

The four standard termination notions are then separated by whether they quantify over one input or all inputs, and whether they ask only for probability-$1$ termination or for finite expected runtime (Kaminski et al., 2015).

Notion	Condition	Universality
AST	$\Pr_{P,\eta}(\downarrow)=1$	Fixed input
UAST	$\forall \eta:\ (P,\eta)\in AST$	All inputs
PAST	$E_{P,\eta}(\downarrow)<\infty$	Fixed input
UPAST	$1$0	All inputs

The inclusion

$1$1

is immediate from finiteness of expected runtime (Kaminski et al., 2015).

The quantified object varies with the operational model. In probabilistic control-flow graphs with demonic nondeterminism, “positive termination” already means that under each scheduler $1$2 and for each initial valuation $1$3, $1$4, so the universal quantification is built into the base notion (Agrawal et al., 2017). For probabilistic pushdown automata, PAST is defined universally over all initial states $1$5 by requiring $1$6 (Winkler et al., 2023). For nondeterministic recursive probabilistic programs, the corresponding notion is called bounded termination and is defined by

$1$7

from a stack element $1$8 (Chatterjee et al., 2017). In probabilistic term rewriting, the closest analogue is full PAST (fPAST), which requires finite expected derivation length for every full rewrite sequence (Kassing et al., 2023). This suggests that UPAST is best understood as a family of universally quantified finite-expectation properties indexed by the underlying semantic model.

2. Arithmetical complexity and undecidability

For fully probabilistic imperative while-programs, the exact arithmetical-hierarchy classification is known. The central result is Theorem 10 of Kaminski et al.: $1$9 The same paper also places the neighboring problems at lower levels: $P$0

$P$1

The logical reason for the jump is the quantifier pattern

$P$2

which adds a leading universal quantifier over inputs to the $P$3 characterization of PAST (Kaminski et al., 2015).

Outside this first-order while-program setting, exact UPAST complexity classifications are rarer. For higher-order probabilistic computation, the corresponding qualitative AST problem is already undecidable and $P$4-complete in the general background considered by recent PHORS work (Lago et al., 30 Apr 2026). In probabilistic lambda-calculi, AST and PAST inherit $P$5- and $P$6-hardness phenomena via encodings of probabilistic Turing machines, and the resulting type-theoretic characterizations are exact only as suprema over derivations rather than as finite certificates (Lago et al., 2020). This suggests that the explicit $P$7-completeness theorem for UPAST remains one of the sharpest general complexity results presently available.

3. UPAST versus qualitative almost-sure termination

A persistent misconception is that almost-sure termination is “nearly” the same as positive almost-sure termination. The literature repeatedly separates the two. The symmetric random walk with constant step size is the standard counterexample: it is AST but not PAST (Winkler et al., 28 Apr 2025). In the LexRSM literature, there is an explicit example of a program admitting a 2-dimensional linear LexRSM while still having infinite expected termination time, so general LexRSM existence implies universal AST but not UPAST (Agrawal et al., 2017). In the higher-order lambda-calculus setting, the paper on non-idempotent intersection types likewise presents AST terms whose expected runtime is infinite, showing that the semantic gap persists even in exact type-theoretic characterizations (Lago et al., 2020).

This distinction explains why many proof methods are intentionally qualitative. Terminating patterns prove that a set of runs of probability $P$8 is terminating, but they do not analyze expected runtime (Esparza et al., 2012). Guarded refinement for higher-order programs with state proves lower bounds on termination probability by relating a program to a simpler Markov model, but it does not define or transfer expected-time bounds (Gregersen et al., 2024). The pGCL rule based on a real-valued supermartingale with parametric progress functions $P$9 and $\eta$0 is expressly designed to cover AST examples such as random walks with infinite expected time, not positive termination (McIver et al., 2017). Descent supermartingales provide a modular AST rule for probabilistic programs, but their soundness theorem concludes only $\eta$1 for all schedulers, not $\eta$2 (Huang et al., 2019).

Accordingly, UPAST should not be viewed as a routine strengthening of UAST. It requires proof principles that control the expectation of the termination time, not merely the measure of nonterminating runs.

4. Proof principles for universal finite expectation

For nondeterministic recursive probabilistic programs, the strongest general result is that ranking measure functions are sound and complete for bounded termination. A ranking measure function $\eta$3 assigns a nonnegative value to each stack element and satisfies one-step decrease conditions for assignments, calls, branches, and nondeterministic choices. The soundness theorem yields

$\eta$4

where

$\eta$5

Completeness is equally sharp: the function $\eta$6 itself is a ranking measure function with $\eta$7 (Chatterjee et al., 2017). In this model, the ranking-supermartingale approach is therefore not merely sufficient but exact for the universal finite-expectation property.

For probabilistic pushdown automata, the corresponding universal PAST problem is characterized by certificates. The expected runtimes are the least solution of

$\eta$8

and PAST holds iff there exist rational vectors $\eta$9 such that

$\Pr_{P,\eta}(\downarrow)$0

These certificates are polynomially checkable, and for the pBPA subclass the paper gives a PTIME decision result (Winkler et al., 2023).

Lexicographic ranking supermartingales occupy an intermediate position. In general they prove only universal AST. However, if a strict LexRSM additionally satisfies bounded expected conditional increase, then finite expected termination time follows, with

$\Pr_{P,\eta}(\downarrow)$1

At the program level this yields $\Pr_{P,\eta}(\downarrow)$2 for each scheduler and each initial valuation under the bounded-ECI assumptions (Agrawal et al., 2017). This is not a complete UPAST method, but it is a direct route from a scheduler-uniform martingale certificate to universal finite expectation.

5. Restricted decidable fragments and model-specific reductions

In higher-order probabilistic recursion schemes, bounded fragments admit exact symbolic reductions of AST and PAST to algebraic generating functions. For a PHORS $\Pr_{P,\eta}(\downarrow)$3, the generating function

$\Pr_{P,\eta}(\downarrow)$4

satisfies

$\Pr_{P,\eta}(\downarrow)$5

Using bounded exponentials in the weighted relational model of linear logic, the paper proves decidability of AST and PAST for finitely bounded and related bounded fragments, because the associated generating functions are algebraic (Lago et al., 30 Apr 2026). This is a fragment result rather than a general UPAST theorem, but it is one of the clearest exact higher-order decision procedures for positive termination.

For probabilistic term rewriting, the relevant universal notion is fPAST. The central transfer theorems show that, under appropriate structural restrictions, innermost positive termination suffices for full positive termination. In particular, for non-overlapping linear PTRSs,

$\Pr_{P,\eta}(\downarrow)$6

and for non-overlapping right-linear systems,

$\Pr_{P,\eta}(\downarrow)$7

Orthogonal spare systems admit analogous equivalences on basic terms (Kassing et al., 2023). The earlier paper on probabilistic term rewriting proves that interpretation methods are sound and complete for SPAST, a strengthening of PAST requiring a uniform finite upper bound on expected derivation length over all reductions from a term; SPAST therefore implies UPAST/fPAST but does not coincide with it (Avanzini et al., 2018).

A specialized but notable quantitative result concerns polynomial random walks. For the loop

$\Pr_{P,\eta}(\downarrow)$8

PAST holds when the step-size degree dominates both the drift degree and a probability-dependent threshold: $\Pr_{P,\eta}(\downarrow)$9 The proof proceeds by inductive bounds on the cumulative distribution function and survival probabilities rather than by martingales or invariants (Winkler et al., 28 Apr 2025). This shows that UPAST-style finite-expectation reasoning can also arise from distributional tail analysis instead of ranking-based certificates.

6. Automation, applications, and scope boundaries

Automation is uneven across models. Amber handles a restricted class of probabilistic while-programs with polynomial arithmetic and supports both AST and PAST analysis. Its terminology already treats AST as termination with probability $E_{P,\eta}(\downarrow).$0 on all inputs and PAST as finite expected runtime; it also supports symbolic constants, so a successful proof can serve as a parameter-uniform certificate over the assumed symbolic domain (Moosbrugger et al., 2021). In probabilistic term rewriting, AProVE implements the innermost-to-full transfer criteria, allowing full AST and related properties to be established by reducing them to innermost analyses when the syntactic hypotheses apply (Kassing et al., 2023).

At the same time, the boundary between UPAST and AST-only methods remains sharp. Pattern-based refinement loops, guarded refinement for higher-order state, generalized LexRSMs, descent supermartingales, and pGCL variant rules each show that powerful universal AST methods can succeed far beyond the reach of finite-expectation arguments (Esparza et al., 2012). Conversely, certificate-based pPDA analysis, ranking measure functions for recursive probabilistic programs, and fragment-specific algebraic or distributional methods show that universal finite expectation can sometimes be captured exactly, but usually only with stronger structural assumptions or more specialized semantics (Winkler et al., 2023).

Taken together, the literature presents UPAST as a genuinely stronger liveness property rather than a minor quantitative refinement. In first-order imperative while-programs it is one full arithmetical-hierarchy level harder than PAST and strictly above AST and UAST (Kaminski et al., 2015). In richer models, the property persists under different names—bounded termination, positive termination, full PAST, or universally quantified expected runtime—but the central challenge is the same: proving that every admissible execution mode not only terminates with probability $E_{P,\eta}(\downarrow).$1, but does so with finite expectation.