Discrete Minimum-Step Distance

• Discrete minimum-step distance is a metric that quantifies the minimum hops or actions between configurations in systems like quantum mechanics, MDPs, and directed graphs.
• It underpins computational techniques including BFS, Floyd–Warshall, and embedding methods (e.g., MadDist/TDMadDist) to efficiently approximate state-to-state transitions.
• Recent advancements offer rigorous approximation algorithms for DAGs and quantum path integrals, bridging discrete combinatorial models with continuous phenomena.

Discrete minimum-step distance is a foundational metric concept with manifestations across quantum mechanics, Markov decision processes (MDPs), and the structural analysis of directed graphs. It quantifies the least number of discrete “hops” or “actions” separating two configurations—be they spatial positions, process states, or graph nodes. This metric underpins combinatorial path-summing frameworks, state embedding algorithms, and computational geometry, bridging ideas from statistical mechanics, dynamic programming, and graph theory.

1. Formal Definitions and Metric Properties

In the discrete setting, minimum-step distance (sometimes called minimum action distance, MAD) is defined as the minimal length of a valid sequence that transitions from one element to another under a specified local move rule. Given a set of states $\mathcal S$ and a one-step relation $R \subseteq \mathcal S \times \mathcal S$, the minimum-step distance between $s,s'\in\mathcal S$ is

$$d_{\mathrm{min}}(s,s') = \min\{ K \mid \exists (s_0, ..., s_K): s_0=s,\, s_K=s',\, (s_k,s_{k+1})\in R \}$$

In Markovian systems with transition kernel $P$, $R = \{(s,s') \mid \exists\, a: P(s'|s,a)>0\}$ gives the action-induced connectivity, and $d_{\mathrm{MAD}}$ is the minimal action count connecting $s$ to $s'$ (Steccanella et al., 10 Jun 2025).

For directed graphs, including DAGs, letting $d_G(u,v)$ be the shortest path from $R \subseteq \mathcal S \times \mathcal S$0 to $R \subseteq \mathcal S \times \mathcal S$1, the undirected minimum-step metric is

$R \subseteq \mathcal S \times \mathcal S$2

which is symmetric and satisfies nonnegativity, identity, and the triangle inequality (Dalirrooyfard et al., 2021).

In the quantum context, the minimum-space-interval hypothesis postulates a discrete spatial step $R \subseteq \mathcal S \times \mathcal S$3, and classical paths correspond to sequences with the minimal step count connecting initial and final positions (Ajaib, 2014).

2. Discrete Minimum-Step Distance in Quantum Path Integrals

Path integral formulations of quantum mechanics admit a discrete minimum-step structure wherein the state of a particle transitions between positions in units of $R \subseteq \mathcal S \times \mathcal S$4. For a move from $R \subseteq \mathcal S \times \mathcal S$5 to $R \subseteq \mathcal S \times \mathcal S$6 ($R \subseteq \mathcal S \times \mathcal S$7), the minimal path contains $R \subseteq \mathcal S \times \mathcal S$8 steps (all forward), with nonminimal paths taking $R \subseteq \mathcal S \times \mathcal S$9 steps, where $s,s'\in\mathcal S$0 counts backward-forward flips.

The combinatorial weight for each such path is

$s,s'\in\mathcal S$1

In the Euclidean propagator, all paths of length $s,s'\in\mathcal S$2 contribute with action $s,s'\in\mathcal S$3, leading to the sum-over-histories propagator:

$s,s'\in\mathcal S$4

with $s,s'\in\mathcal S$5 (Ajaib, 2014).

When the step size $s,s'\in\mathcal S$6 exceeds the de Broglie wavelength $s,s'\in\mathcal S$7, only the minimal step path significantly contributes, recovering the continuous limit under $s,s'\in\mathcal S$8.

3. Minimum Action Distance in Markov Decision Processes

In deterministic and stochastic MDPs, the minimum action distance between states $s,s'\in\mathcal S$9 and $$d_{\mathrm{min}}(s,s') = \min\{ K \mid \exists (s_0, ..., s_K): s_0=s,\, s_K=s',\, (s_k,s_{k+1})\in R \}$$0 is the length of the shortest feasible state-only trajectory, regardless of reward signals or executed actions. For finite $$d_{\mathrm{min}}(s,s') = \min\{ K \mid \exists (s_0, ..., s_K): s_0=s,\, s_K=s',\, (s_k,s_{k+1})\in R \}$$1, constructing the digraph $$d_{\mathrm{min}}(s,s') = \min\{ K \mid \exists (s_0, ..., s_K): s_0=s,\, s_K=s',\, (s_k,s_{k+1})\in R \}$$2 (unit-cost edges) allows computing all-pairs minimum-step distances by:

• Floyd–Warshall: $$d_{\mathrm{min}}(s,s') = \min\{ K \mid \exists (s_0, ..., s_K): s_0=s,\, s_K=s',\, (s_k,s_{k+1})\in R \}$$3 for exact all-pairs solution
• BFS from a single source: $$d_{\mathrm{min}}(s,s') = \min\{ K \mid \exists (s_0, ..., s_K): s_0=s,\, s_K=s',\, (s_k,s_{k+1})\in R \}$$4, exact for a single state

For large or continuous state spaces, embedding methods seek to learn $$d_{\mathrm{min}}(s,s') = \min\{ K \mid \exists (s_0, ..., s_K): s_0=s,\, s_K=s',\, (s_k,s_{k+1})\in R \}$$5 so that $$d_{\mathrm{min}}(s,s') = \min\{ K \mid \exists (s_0, ..., s_K): s_0=s,\, s_K=s',\, (s_k,s_{k+1})\in R \}$$6 (Steccanella et al., 10 Jun 2025).

4. Approximation Algorithms for Min-Distance in DAGs

Dalirrooyfard and Kaufmann (Dalirrooyfard et al., 2021) establish approximation algorithms for the minimum-step metric in DAGs, particularly for min-radius ($$d_{\mathrm{min}}(s,s') = \min\{ K \mid \exists (s_0, ..., s_K): s_0=s,\, s_K=s',\, (s_k,s_{k+1})\in R \}$$7) and min-diameter ($$d_{\mathrm{min}}(s,s') = \min\{ K \mid \exists (s_0, ..., s_K): s_0=s,\, s_K=s',\, (s_k,s_{k+1})\in R \}$$8). For the min-radius:

• A $$d_{\mathrm{min}}(s,s') = \min\{ K \mid \exists (s_0, ..., s_K): s_0=s,\, s_K=s',\, (s_k,s_{k+1})\in R \}$$9-approximation is computed in $P$0 time via block partitioning, multi-source BFS, and binary search.
• For sparse DAGs, any algorithm achieving $P$1-approximation requires $P$2 time (Hitting Set Conjecture), establishing conditional tightness.

For the min-diameter, a $P$3-approximation is achievable in dense DAGs in $P$4 time, relying on hitting set arguments and Boolean matrix multiplication. Hardness results stemming from the Orthogonal Vectors Conjecture indicate these bounds are unlikely to be surpassed under standard complexity-theoretic assumptions.

5. Learning Minimum-Step Embeddings from Trajectories

MadDist and TDMadDist frameworks (Steccanella et al., 10 Jun 2025) learn embeddings that accurately reflect minimum-step distances from state-only trajectories. The approach samples pairs $P$5 from trajectories and applies:

• Scaled regression loss ensuring $P$6
• Contrastive separation loss to guarantee sufficient spread among unrelated pairs
• Constraint loss to enforce that short-horizon pairs do not overestimate $P$7

In TDMadDist, a target network introduces a bootstrap term akin to temporal-difference learning. Both methods accommodate symmetric and asymmetric (quasimetric) distance functions, with the ReLU-based "simple" quasimetric satisfying necessary metric properties. Empirical results on grid-world benchmarks demonstrate Pearson $P$8 and coefficient of variation $P$9, surpassing prior embedding methods.

6. Statistical and Ensemble Interpretations

Discrete minimum-step distances are structurally analogous to statistical mechanical ensembles. In the 1D quantum path context, paths correspond to configurations of $R = \{(s,s') \mid \exists\, a: P(s'|s,a)>0\}$0 two-level spins constrained to a target magnetization. The partition function encodes path multiplicities, with the sum over paths paralleling the sum over spin configurations. In the classical limit ($R = \{(s,s') \mid \exists\, a: P(s'|s,a)>0\}$1), only the minimal step (all spins aligned) path survives—the discrete minimum-step path (Ajaib, 2014).

7. Practical Recommendations and Connections

• For small discrete systems, brute-force computation of minimum-step distance via BFS or Floyd–Warshall is tractable.
• In settings with large, continuous, or partially observed state spaces, embedding-based learning with MadDist/TDMadDist gives practical and high-fidelity approximations (Steccanella et al., 10 Jun 2025).
• Choice of distance function (metric vs. quasimetric) should reflect task asymmetry—Euclidean for symmetric distances, ReLU-based for directional tasks.
• The minimum-step hypothesis forms a bridge between discrete combinatorial models and continuous propagation, critically depending on the system’s granulometry (e.g., spatial step vs. de Broglie wavelength).
• Approximation complexities for the minimum-step metric in DAGs are tightly constrained by computational hardness, guiding both algorithm design and expectations about tractable precision (Dalirrooyfard et al., 2021).

Discrete minimum-step distance is thus a unifying metric, operating across classical and quantum systems, algorithmic graph theory, and the analysis and learning of dynamical environments.