Minimum Strategy in Safety Graph Games

• Minimum Strategy denotes the smallest finite-memory needed for a winning strategy in graph games, measured by the antichain width of residual safety languages.
• The approach leverages order-theoretic techniques by analyzing left quotients and their inclusion order to precisely characterize memory needs.
• Constructive synthesis maps minimal winning residuals to memory states, providing a structural method to build optimal controllers even beyond regular objectives.

Searching arXiv for the focal paper and closely related work on memory in graph games. “Minimum strategy” in the context of graph games with safety objectives denotes the least amount of finite memory required by a winning player to ensure that no forbidden finite prefix is ever produced along a play. In two-player games on colored directed graphs, the paper “Playing Safe, Ten Years Later” identifies this quantity exactly for arbitrary safety objectives by relating it to the order-theoretic structure of the objective’s left quotients (Colcombet et al., 2022). The central result is that the minimum number of memory states needed by a winning strategy is the width of the partially ordered set of residual safety languages under inclusion; this holds on finite-degree arenas without regularity assumptions, and on arbitrary arenas when the quotient order is well-founded (Colcombet et al., 2022).

1. Formal setting and the notion of minimum strategy

The game model is a two-player game over a colored arena

$A=(V,(V_E,V_A),E,c),$

where $V_E$ and $V_A$ are Eve’s and Adam’s vertices, $E\subseteq V\times V$ is the edge relation, and $c:E\to\mathbb{A}$ labels edges over a finite alphabet $\mathbb{A}$ (Colcombet et al., 2022). A play is an infinite path whose induced color sequence lies in $\mathbb{A}^\omega$ (Colcombet et al., 2022).

A safety objective is specified by a set $P\subseteq \mathbb{A}^*$ of forbidden finite prefixes. The induced safety language is

$$Safe(P)=\{\rho=a_0a_1a_2\cdots\in \mathbb{A}^\omega \mid \forall i,\; a_0a_1\cdots a_i\notin P\}.$$

Eve wins if she can ensure that no prefix of the color sequence ever belongs to $P$ (Colcombet et al., 2022).

A strategy for Eve maps finite histories ending in an Eve vertex to legal outgoing edges. Finite-memory strategies are implemented by a memory structure

$V_E$0

with memory states $V_E$1, initial state $V_E$2, and update function $V_E$3, together with a next-move function $V_E$4 (Colcombet et al., 2022). The question is then intrinsic: among all winning strategies from winning positions, what is the smallest possible cardinality $V_E$5?

The paper formalizes the worst-case minimum memory requirement of a safety objective $V_E$6 by a quantity $V_E$7, defined as the supremum, over all games with objective $V_E$8 and all winning initial vertices, of the minimum memory size of a winning strategy from that vertex (Colcombet et al., 2022). A finite-degree variant $V_E$9 restricts the supremum to arenas of finite degree (Colcombet et al., 2022). This is the precise meaning of a “minimum strategy”: not merely a winning strategy with some finite-memory upper bound, but a strategy using the smallest possible number of memory states in the worst case (Colcombet et al., 2022).

2. Left quotients, inclusion order, and antichain width

The main combinatorial object is the family of left quotients, or residual objectives. For a safety objective $V_A$0 and a finite word $V_A$1,

$V_A$2

This is the remaining objective after the prefix $V_A$3 has already been observed (Colcombet et al., 2022). The set of all such residual languages is denoted $V_A$4 (Colcombet et al., 2022).

These residuals are ordered by language inclusion: $V_A$5 An antichain is a set of pairwise incomparable residuals. The width of $V_A$6 is the size of a largest antichain (Colcombet et al., 2022).

This order-theoretic viewpoint captures how much memory is genuinely needed. If one residual language is included in another, then satisfying the smaller one is stronger; a strategy safe for the smaller residual is automatically safe for the larger one (Colcombet et al., 2022). Hence memory need not distinguish all residuals, only enough incomparable obligations to preserve winning information. Conversely, incomparable residuals represent genuinely different future obligations that the opponent can exploit if they are merged in memory (Colcombet et al., 2022).

The paper’s representative example is the safety objective “never see both $V_A$7 and $V_A$8,” defined by

$V_A$9

After reading $E\subseteq V\times V$0, the residual becomes “only $E\subseteq V\times V$1’s and neutral letters may follow”; after reading $E\subseteq V\times V$2, symmetrically (Colcombet et al., 2022). These two residuals are incomparable under inclusion, so the width is at least $E\subseteq V\times V$3, and two memory states are required to remember which relevant letter was seen first (Colcombet et al., 2022).

3. Exact characterization theorem

The main theorem states: $E\subseteq V\times V$4 for all safety objectives $E\subseteq V\times V$5 (Colcombet et al., 2022). Moreover, if $E\subseteq V\times V$6 is well-founded, then

$E\subseteq V\times V$7

In particular, this applies to regular safety objectives, since regularity implies well-foundedness of the quotient order in the sense used by the paper (Colcombet et al., 2022).

Thus the minimum amount of memory is exactly the size of a largest antichain of left quotients. The result is exact rather than asymptotic, and it applies to arbitrary safety objectives, including non-regular ones, on finite-degree arenas (Colcombet et al., 2022).

The theorem sharpens earlier automata-based upper bounds by replacing the crude estimate $E\subseteq V\times V$8 with the much smaller and exact invariant given by width. It also reframes the memory problem as an order-theoretic one: the complexity of a safety objective is measured not by the total number of residuals but by how many of them are mutually incomparable (Colcombet et al., 2022).

A further corollary identifies the half-positional case. Over finite-degree arenas,

$E\subseteq V\times V$9

So positionality for safety objectives is equivalent to residual languages being totally ordered by inclusion (Colcombet et al., 2022).

4. Why antichain width determines memory

The upper-bound argument begins with the observation that one can always store the current residual objective $c:E\to\mathbb{A}$0, where $c:E\to\mathbb{A}$1 is the play prefix. This already yields $c:E\to\mathbb{A}$2 memory states, but it is generally non-optimal (Colcombet et al., 2022).

In the well-founded case, one considers at each vertex the minimal winning residuals under inclusion. These minimal winning residuals form an antichain, so there are at most

$c:E\to\mathbb{A}$3

such obligations (Colcombet et al., 2022). Memory state $c:E\to\mathbb{A}$4 tracks one minimal obligation, and the strategy preserves the invariant

$c:E\to\mathbb{A}$5

along the current play prefix $c:E\to\mathbb{A}$6 (Colcombet et al., 2022). Since the tracked residual is always a nonempty under-approximation of the actual residual objective, safety is maintained.

When the quotient order is not well-founded, minimal residuals may fail to exist. The paper then uses a Dilworth-style chain decomposition: the set of winning quotients at a vertex is covered by at most $c:E\to\mathbb{A}$7 chains, and memory tracks a chain rather than a single quotient (Colcombet et al., 2022). The invariant becomes that some quotient in the tracked chain under-approximates the current residual. This yields the upper bound on finite-degree arenas without any regularity hypothesis (Colcombet et al., 2022).

The lower bound is dual. Choose a largest antichain

$c:E\to\mathbb{A}$8

For every pair $c:E\to\mathbb{A}$9, incomparability yields a witness $\mathbb{A}$0 belonging to one residual but not the other (Colcombet et al., 2022). The authors then construct a game where Adam first selects which prefix $\mathbb{A}$1 has effectively occurred, Eve must react from her current memory state, and Adam finally uses a witness $\mathbb{A}$2 to expose any incorrect merger of incomparable obligations (Colcombet et al., 2022). If Eve had fewer than $\mathbb{A}$3 memory states, two incomparable residuals would necessarily be identified, and Adam could force a loss. Hence at least $\mathbb{A}$4 states are necessary (Colcombet et al., 2022).

The width therefore yields both the optimal upper bound and the matching lower bound. This is the substantive content of the minimum-strategy theorem (Colcombet et al., 2022).

5. Constructive synthesis of memory-optimal strategies

The result is constructive. In the regular or well-founded case, the proof yields an explicit synthesis method (Colcombet et al., 2022):

1. compute or identify the winning left quotients at each vertex;
2. select the minimal winning quotients under inclusion;
3. index these by memory states $\mathbb{A}$5, where $\mathbb{A}$6 is the width;
4. update memory after reading a color $\mathbb{A}$7 by moving from a quotient $\mathbb{A}$8 to one below $\mathbb{A}$9;
5. at Eve vertices, choose an edge preserving winningness of the tracked residual.

In finite-degree arenas without well-foundedness, the construction replaces antichains of minimal quotients by a cover of the winning quotients with $\mathbb{A}^\omega$0 decreasing chains, and the memory update moves between these chains as the residual objective evolves (Colcombet et al., 2022).

This synthesis aspect is significant because the characterization is not merely existential. The quotient poset does not only determine the optimum value of memory; it also provides the structural template for an optimal finite-memory controller (Colcombet et al., 2022).

A plausible implication is that for non-regular safety objectives, order-theoretic analysis can play the role that automata minimization plays for regular objectives. The paper presents the quotient-width principle as precisely such a foundational replacement (Colcombet et al., 2022).

6. Applications and examples

Two applications are emphasized.

For generalized reachability viewed dually as a safety objective, let the alphabet be $\mathbb{A}^\omega$1 and define

$\mathbb{A}^\omega$2

This means that at least one distinguished color is never seen (Colcombet et al., 2022). The left quotients correspond to strict subsets of $\mathbb{A}^\omega$3, ordered by inclusion. By Sperner’s theorem, the width is

$\mathbb{A}^\omega$4

Therefore

$\mathbb{A}^\omega$5

and this gives the exact memory requirement for the opponent in generalized reachability games (Colcombet et al., 2022).

For games with counters, the paper studies boundedness objectives $\mathbb{A}^\omega$6, requiring the counter to remain at most $\mathbb{A}^\omega$7 under arbitrary monotone counter actions. The residuals are linearly ordered, so width equals $\mathbb{A}^\omega$8, and the objective is half-positional: $\mathbb{A}^\omega$9 The same reasoning applies to energy objectives with monotone updates; positional strategies suffice because the residual languages are linearly ordered by inclusion (Colcombet et al., 2022).

These examples illustrate two extremes: one family where width grows combinatorially, and another where width collapses to one. In both cases the characterization yields exact memory rather than upper and lower bounds separated by a gap (Colcombet et al., 2022).

7. Scope, caveats, and significance

The theorem has an important qualification. In full generality, exact characterization by width requires either finite-degree arenas or well-foundedness of the quotient order (Colcombet et al., 2022). The paper gives the outbidding objective as a counterexample showing that finite degree cannot be dropped without replacement: although the quotient poset has width $P\subseteq \mathbb{A}^*$0, there exists an infinite-degree arena where Eve can win only with infinite memory (Colcombet et al., 2022). So width is not an unconditional characterization on arbitrary arenas.

This caveat separates two regimes. On finite-degree arenas, width always gives the exact minimum memory for safety objectives. On arbitrary arenas, width remains exact only when the quotient order is well-founded (Colcombet et al., 2022). Regular safety objectives fall into the latter class.

The broader significance of the work lies in its generality. The paper presents the result as a sharpened formulation of a 2014 insight, now framed as a foundational principle: the exact memory requirement of a safety objective is the width of its residual-language poset (Colcombet et al., 2022). It also notes that this result became a stepping stone for later developments on positionality, chromatic memory, universal graphs, and memory characterizations beyond regular objectives (Colcombet et al., 2022).

At the same time, the paper is explicit about its limits. The characterization is precise for safety objectives, including non-regular ones, but the wider program of characterizing memory requirements for all $P\subseteq \mathbb{A}^*$1-regular objectives remains open (Colcombet et al., 2022). A reasonable interpretation is that safety provides a model case in which topology and order theory fully determine memory, while more general winning conditions still resist such an exact description.

In summary, the paper’s answer to the minimum-strategy problem is exact and structural: $P\subseteq \mathbb{A}^*$2 For safety games, memory is neither governed by automaton size alone nor by ad hoc strategy constructions, but by the antichain complexity of the residual objective space (Colcombet et al., 2022).