Directed Latency: Constant-Factor Approximation

• The paper introduces the first polynomial-time constant-factor approximation for Directed Latency using innovative LP strengthening and a novel bucketing framework.
• It employs a rank-based bucketing method with limited guessing to partition clients efficiently and control additional latency from tour stitching.
• The approach overcomes the long-standing O(log n) approximation barrier, offering promising extensions to asymmetric routing and related optimization problems.

The Directed Latency problem concerns optimizing the traversal of an asymmetric metric space by a single vehicle (or agent), originating at a depot, such that the total latency—the sum of waiting times of all clients before they are visited—is minimized. Key challenges stem from the asymmetric nature of distances, the lack of tour symmetry, and the combinatorial complexity of the optimal visitation order. Historically, while the undirected variant (the Deliveryperson or Traveling-Repairperson problem) has long enjoyed constant-factor approximations, the best known bounds for the directed case remained $O(\log n)$-approximations for more than a decade. Recent breakthroughs have yielded the first constant-factor approximations: initially in quasi-polynomial time, and now—through the development of innovative LP-based techniques and a new bucketing framework—in truly polynomial time (Friggstad et al., 2019, Blauth et al., 17 Dec 2025).

1. Problem Formulation and Key Definitions

The Directed Latency problem is defined on a directed complete graph $(V \cup \{r\}, c)$ where $c$ is an asymmetric metric: $c(u, v) \geq 0$, $c(u, u)=0$, and $c(u, v) \leq c(u, w) + c(w, v)$ for all $u, v, w$. The depot or root is denoted $r$. A solution is an $r$-rooted Hamiltonian path $P$ visiting all $v \in V$ exactly once. The latency $c_P(v)$ of a node $v$ is its distance from the depot along $P$, with total latency $\ell(P) = \sum_{v \in V} c_P(v)$. The objective is to compute a path $P$ minimizing $\ell(P)$.

This formulation naturally generalizes the undirected Deliveryperson problem to asymmetric settings. In the undirected case, the minimum-latency path is more tractable; for asymmetric metrics, structural asymmetries radically increase hardness, both algorithmically and integrality-wise.

2. Prior Results and the $O(\log n)$-Barrier

Friggstad, Salavatipour, and Svitkina [FS13] introduced an $O(\log n)$-approximation using geometrically growing "buckets" of the time horizon: clients are grouped according to their optimum latency intervals $I_i=[2^{i-1}, 2^i)$. In each bucket, an auxiliary (orienteering or path-TSP) subproblem is solved to build short subpaths, which are then concatenated. Each stitching step adds only $O(2^i)$, but summing over all levels accumulates to the logarithmic overhead.

Subsequently, in (Friggstad et al., 2019), a quasi-polynomial-time constant-factor approximation was achieved by augmenting the natural LP relaxation with "guesses"—explicit enforcement that a representative client per bucket is visited at a specific time. This reduces the blowup of the patching process, but as the number of buckets is $O(\log n)$, the overall runtime grows to $n^{O(\log n)}$.

A pivotal limitation of all previous techniques was their inability to enforce tight enough LP constraints globally, especially when controlling the ordering of visits between far-apart nodes without an inordinate enumeration of possibilities.

3. Polynomial-Time Constant-Factor Algorithm: New Bucketing and LP Strengthening

The first polynomial-time constant-factor approximation for Directed Latency is achieved by fundamentally reworking the bucketing/guessing approach (Blauth et al., 17 Dec 2025). Key innovations include:

• Bucketing by Rank, Not Value: Clients are grouped by their position (rank) in the optimal sequence rather than by latency magnitude. For $|V|+1=2^k$, bucket $B^*_1$ contains the first $(n+1)/2$ clients, $B^*_2$ the next $(n+1)/4$, etc. Buckets are merged as dictated by latency thresholds $t_i$, yielding at most $O(\log n)$ groups $G^*_1,\dots,G^*_q$, where $t_{i+1} \geq \frac{4}{3} t_i$.
• Tour-Intervals and Roots: Each group $G^*_i$ is classified as a "tour-interval" if the optimal path $P^*$ contains a short back-edge into $G^*_i$, forming a cycle with controlled length. For each tour-interval, a single "root" client $r_i$ is identified such that all clients in $G^*_i$ are accessible within $O(t_{i+1})$ asymmetric distance from $r_i$.
• Limited Guessing: Instead of guessing one vertex per bucket (which leads to quasi-polynomial enumeration), only the tour-intervals and their roots are guessed—now a polynomial number.
• Strengthened LP: The time-indexed LP is formulated on a layered, time-expanded graph $G_T$ using the variables $x_{v,t}$ (visit time) and $z_{(u,t)\to(v,t')}$ (transition). Two forms of structural constraints are added:
  • Flows can only enter or leave a tour-interval at its root.
  • For pairs $(u,v)$ with $c(u,v)$ not $t_{i+1}$-short, explicit ordering constraints are enforced, ensuring that either $v$ precedes $u$ or the ordering can be repaired via inclusion in a tour-cycle.

Although the resulting LP is no longer a relaxation (i.e., the fractional optimum may not coincide with the true optimum of Directed Latency), it can still be solved in polynomial time, and a feasible fractional solution with cost $O(1)\, \mathrm{OPT}$ exists.

4. Rounding Techniques and Composition of the Final Solution

The rounding procedure operates in stages:

• Partitioning: Clients are split into "tour-clients" (substantial mass within tour-intervals) and "non-tour-clients."
• Tour-Clients: For each tour-interval, a fractional circulation on its induced subgraph is transformed (via splitting-off and LP-relative ATSP-path solutions) into a closed tour covering the assigned tour-clients at controlled cost.
• Non-Tour-Clients: Buckets are created in non-tour intervals by mass aggregation. Clients with "long" pairwise distances to earlier buckets are absorbed as necessary, preserving cut requirements. For each such bucket, an $s$-to-sink path (via the LP-relative ATSP-path algorithm at connectivity parameter $\rho$) is constructed, whose cost is again proportional to the corresponding threshold $t_i$.
• Stitching: The tours and paths are sequentially concatenated. Each joining edge incurs at most $O(t_i)$ cost due to geometric progression in thresholds and the structure of $z$-flow. The telescoping structure controls the total additional latency, preserving the constant-factor bound.

5. Analytical Guarantees and Complexity

The algorithm guarantees, for an explicit constant $C$,

$$\ell(P) \leq C \sum_{v,t} ub(t) x_{v,t} \leq C \cdot O(1) \cdot \mathrm{OPT}$$

where $ub(t)$ is a calibrated upper bound related to the bucket intervals. The number of configurations—choices of thresholds, tour-intervals, and roots—is polynomial in $n$. All LPs and subroutines are polynomial-time (min-cut separation for flows, ATSP-path rounding, splitting-off for circulations). The constant $C$ is not fully optimized; its magnitude depends on structural properties of the LP and rounding procedures, and is on the order of $10^7$ in the presented formulation.

The removal of exhaustive guessing over all buckets—crucial in previous quasi-polynomial-time schemes—now enables the solution to run in polynomial time without sacrificing approximation quality (Blauth et al., 17 Dec 2025).

6. Extensions, Special Cases, and Broader Implications

The bucketing and constraint-driven rounding framework is sufficiently general to adapt to related latency objectives. For regret-minimization or time-window constraints (where the goal is, e.g., to minimize excess waiting time relative to an ideal direct route), the combination of tour-interval identification and flow-based LP strengthening can be adapted, provided analogous certificates for the existence of short cycles (and thus roots) can be constructed for subsets of clients.

For symmetric metrics equipped with regret objectives, the approach yields improved approximation constants via the analysis of the regret metric, and a careful adaptation of the ATSP-path subroutines further tightens the integrality gap (Friggstad et al., 2019).

These advances suggest a broader impact on the theory of approximation algorithms for asymmetric vehicle routing and combinatorial search, introducing structural tools (such as root-constrained tour intervals and constrained flow models) that are likely to influence further developments.

7. Open Questions and Current Limitations

Despite the polynomial-time constant-factor approximation, several limitations and avenues for improvement remain:

• The explicit constant in the approximation, while theoretically finite, is large; further optimization and tightening of the tour-path rounding process could substantially decrease its practical impact.
• The LP used is not a true relaxation, raising subtle questions about integrality gaps and possible structural weakness at boundaries of the feasible region.
• Extending these techniques to variants with multiple vehicles, additional side constraints (e.g., time windows, precedence), or adapting the framework to more general asymmetric metrics remain open and substantial challenges.
• The conjectured existence of substantially smaller constants, or improved integrality gap bounds for the key subproblems (notably, the ATSP-path LP with relaxed cuts), represents an active research direction.

The emergence of polynomial-time constant-factor algorithms marks a resolution to a longstanding open problem in approximation, but also reorients focus to fine-grained structural control and algorithmic generalization in the asymmetric domain (Friggstad et al., 2019, Blauth et al., 17 Dec 2025).