Gas pipeline operators routinely ask timing questions of their network: does the pressure at a junction stay above a minimum, except for dips shorter than a few seconds? Does a pressure wave arriving from one branch clear within a set time after a valve closes? Today these questions are answered by running a simulation and reading the answer off the computed trace. Nobody states how wrong that answer might be. The most error-prone place in the simulation is the junction itself. There, the physically correct rule, called equal pressure coupling (the gas pressure must be continuous across the node even when the connected pipes carry gas at different temperatures and therefore different sound speeds), is an implicit nonlinear condition that must be solved at every time step. We propose to solve the network as a fixed-point iteration whose convergence rate is certified by the contraction mapping principle, the theorem that an iteration shrinking distances by a factor $k<1$ converges to a unique answer and comes with a computable error bound at every step. That error bound turns into a guaranteed band around the junction pressure trace. We then check the timing specification on the band rather than on a single trace, and the band is enough to return a verdict of "satisfied", "violated", or "not yet decided".
The link to temporal logic is endpoint decomposition. This technique represents every time interval by its start and end instants, so that questions about intervals become linear inequalities between points. A specification such as "no dip below $p_{\min}$ lasts longer than $D$" concerns the interval during which the predicate $p(t) < p_{\min}$ holds. Under an error band, each endpoint of that interval is known only to lie inside a bracket. The metric requirement then becomes a small system of difference constraints on bracketed points, which can be solved in polynomial time. The claim is that contraction bounds, endpoint brackets, and point-based constraint checking together give sound, anytime verdicts for timed specifications on hyperbolic gas networks with equal pressure junctions. Such verdicts are not available from direct simulation or from the monotone abstractions used for traffic.
What exists and what is missing
Three lines of prior work each cover part of this, and none covers it whole.
The coupling line supplies the physics and the test problems. Coupling conditions for gas networks governed by the isothermal Euler equations compared a simpler subsonic outflow condition against equal pressure coupling on pipes with sound speeds 6 and 8. It showed that only equal pressure avoids a non-physical pressure jump at the node, but it reports no size for the jump and assumes forward flow only. An Asymptotic-Preserving Scheme for Isentropic Flow in Pipe Networks gives a semi-implicit scheme that keeps pressure-continuity coupling well-posed at low Mach number. A central scheme for coupled hyperbolic systems avoids nonlinear interface solves through relaxation. We reuse these schemes and test cases as the pipe-interior solvers. None of them attaches an error guarantee to the junction trace, or asks timing questions of it.
The contraction line supplies the method. A macroscopic traffic flow model with finite buffers on networks proved well-posedness of junction dynamics for the scalar traffic model by showing that a queue-update map contracts over short time windows. That is the template we follow, but we move it from a single-equation traffic model to the two-equation Euler system with equal pressure. Flow and Elastic Networks on the n-torus builds a contraction iteration for static network equilibria and reports 4x to 12x speedups over MATLAB solvers on the IEEE 24-bus case. We borrow its point that a contraction gives guarantees that Newton's method does not. Automatic differentiation of nonsmooth iterative algorithms constrains our design: its finding that momentum-accelerated iterations can have unstable derivatives is a reason to keep plain Picard iteration (repeated application of the map), so that downstream sensitivity studies stay possible.
The temporal line supplies the checking machinery. Constraint Propagation Algorithms for Temporal Reasoning established that point-based constraint reasoning is cubic-time while full interval algebra is NP-hard. Traffic Network Control From Temporal Logic Specifications enforces specifications written in linear temporal logic (a formal language for statements about how a system behaves over time) on traffic networks. It does so through box abstractions that are valid only because the queue dynamics are monotone, meaning that raising one state never lowers another. Acoustic reflections at gas junctions break that monotonicity. Data Complexity and Rewritability of Ontology-Mediated Queries in Metric Temporal Logic under the Event-Based Semantics leaves open how to reason over continuous states rather than isolated events. Our endpoint brackets address a narrow, concrete piece of that gap: they convert a continuous signal known to lie within a band into a finite set of uncertain event points.
Core Claim
We partition the simulation horizon into windows of length $\Delta t$. Within a window, each pipe is solved independently given the time history of pressure and mass flux at its junction ends. The junction then updates those histories by enforcing mass balance and equal pressure. Call this combined update $T$, acting on the vector $x$ of junction boundary histories over the window, measured in the sup norm. For windows short enough that waves reflected at the junction cannot travel to the far pipe end and return, we expect $T$ to contract. The contraction factor $k$ should be governed by the products of acoustic reflection coefficients at the node, which in turn depend on the sound-speed ratios. Picard iteration then gives, after $n$ steps, the standard bound
$$\|x_n - x^*\|_\infty \le \frac{k^n}{1-k}\,\|x_1 - x_0\|_\infty = \varepsilon_n,$$
where $x^*$ is the exact discrete solution of the window. Every junction pressure value is therefore known to within $\varepsilon_n$.
For a threshold predicate, each crossing time is bracketed by the first and last times at which the band straddles the threshold. Near a smooth crossing the bracket width is about $\varepsilon_n / |\dot p|$, from a first-order Taylor expansion. Near a shock arrival the bracket collapses onto the shock arrival time, because the jump is much larger than $\varepsilon_n$. The timed requirement is compiled into difference constraints on the bracketed endpoints and checked twice. If it is satisfied for every point in the brackets, the verdict is "satisfied". If it fails for every point, the verdict is "violated". Otherwise the verdict is "undecided", and we run more iterations. The iteration stops as soon as the verdict is decided, so the cost of verification adapts to how close the system is to the specification boundary.
The certificate is relative to the discretized model. Discretization error is estimated separately, by grid refinement, and is not certified. We say this plainly rather than hide it.
Research Program
The analytical work is to prove contraction of $T$ for the linearized subsonic isothermal system at a junction of arbitrary degree with unequal sound speeds, and to give an explicit $k(\Delta t, a_1,\dots,a_m)$. We then test empirically how far this estimate holds in the nonlinear regime. When shocks collide at the node, contraction can fail. We will detect failure through growth of $\|x_{n+1}-x_n\|$ and respond by halving the window. How often this happens is itself a result worth reporting.
The engineering work is to wrap the asymptotic-preserving and central-scheme pipe solvers as black-box window solvers, implement the junction update, and implement a small endpoint-constraint checker for a fragment of metric temporal logic with bounded windows: "always within $[a,b]$", "eventually within $[a,b]$", and duration bounds on sub-threshold episodes.
Baselines
The naive reference point is direct simulation followed by pointwise checking. We run the same pipe schemes monolithically on the same grid, with the equal pressure junction solved at each step by a bracketed scalar root-finder to machine tolerance, and evaluate the specification on the resulting trace with no error band. This is what practitioners do today. It always gives an answer, but it never says how much to trust that answer. We beat it if our "satisfied" and "violated" verdicts agree with a reference solution, computed at four times the resolution, in essentially every case, while the naive method's wrong verdicts concentrate on near-boundary cases that our method flags as undecided. Compute budgets are matched by wall-clock time. There is no reported result on Lacuna for timed-specification checking on gas networks. For solver cost, the strongest reported figure is that the asymptotic-preserving scheme runs about 140 times faster than an explicit scheme at Mach parameter 0.001 An Asymptotic-Preserving Scheme for Isentropic Flow in Pipe Networks. Our overhead is reported relative to that scheme, not relative to explicit solvers.
The strongest competitor for the certification itself is monolithic Newton with a residual-based error estimate. We solve the full coupled window with Newton, compute the residual, and convert it to a trajectory error using a stability constant estimated by perturbation experiments. We tune it with the same grid and tolerance budget as our method. We beat it if our bound is valid, in the sense that the true error falls inside the band, more often than the residual-based bound, and is no looser than it at equal cost.
The monotone-abstraction competitor follows Traffic Network Control From Temporal Logic Specifications. We bound reachable junction pressures with order-preserving box propagation, which is valid only where the dynamics are monotone, and apply it to the same scenarios. That paper reports a 22.4 second abstraction build on its traffic case study. We expect its bounds to become unsound or vacuous under wave reflection, and we measure where that happens.
Two ablations isolate our ingredients. The first replaces endpoint brackets with sampling: we evaluate the specification on the band at every grid time, which tests whether the point-based compilation matters for decision rate and cost. The second replaces equal pressure with the subsonic condition, which tests whether the choice of coupling changes the verdicts.
Why not just…?
Why not just use the subsonic condition, which is explicit and cheaper? Coupling conditions for gas networks governed by the isothermal Euler equations shows that it produces an artificial pressure jump at the node when sound speeds differ, but it gives no magnitude for the jump. It is therefore an open question whether that jump is large enough to flip threshold verdicts. The coupling ablation settles this. If verdicts rarely differ, the practical case for equal pressure in verification is weak, and we will report that.
Why not just solve the junction with Newton or bisection? At a single node, the equal pressure solve reduces to a monotone scalar mass balance in the common pressure, so bisection is reliable. For that reason we do not claim a contraction advantage for the per-node solve, and the naive baseline uses bisection. The contraction is applied to the network-wide window coupling, where Newton gives no a priori trajectory error bound. Flow and Elastic Networks on the n-torus documents Newton's sensitivity to initialization in cyclic networks. The Newton-plus-residual baseline is the real threat here. If its estimated bounds prove as reliable as ours, then our contribution reduces to the endpoint checker.
Why not just use monotone box abstractions, which already handle temporal logic? They rely on monotonicity, which reflected shocks and choking at equal pressure junctions violate. Our experiments locate the scenarios in which box bounds fail to contain the reference trace.
Why not reason directly with interval relations? Interval-algebra reasoning is NP-hard, whereas point-based reasoning is cubic Constraint Propagation Algorithms for Temporal Reasoning. The sampling ablation measures what the endpoint compilation buys in our own setting.
Proposed Experiments
We use the test cases from the file. The two-pipe interface with sound speeds 6 and 8, and the three-pipe double-reflection row, come from Coupling conditions for gas networks governed by the isothermal Euler equations. The 1-to-2 and 2-to-1 T-junctions, the high-friction inlet-discontinuity case, and the low-Mach sweep come from An Asymptotic-Preserving Scheme for Isentropic Flow in Pipe Networks. On each case we generate families of timed pressure specifications by sweeping the threshold and duration bound across the boundary between satisfied and violated.
We measure five things: verdict soundness against the refined reference; the fraction of cases decided within a fixed iteration budget; wall-clock time to decision; endpoint bracket width at smooth crossings versus shock crossings; and the predicted versus observed contraction factor as the window length and sound-speed ratio vary.
The decisive checks are three. First, the observed $k$ must stay below one, and must track the reflection-coefficient prediction, for windows shorter than a round-trip acoustic travel time. Second, no certified verdict may disagree with the reference, while the naive baseline must err near the specification boundary. Third, the coupling ablation must show whether equal pressure changes verdicts on the variable sound-speed cases. A failure of the first check would mean the certificate is available only in regimes too narrow to matter. A null result on the third check would mean the equal pressure ingredient is not needed for this question.
References
- Coupling conditions for gas networks governed by the isothermal Euler equations
- An Asymptotic-Preserving Scheme for Isentropic Flow in Pipe Networks
- A Central Scheme for Two Coupled Hyperbolic Systems
- A macroscopic traffic flow model with finite buffers on networks: well-posedness by means of Hamilton–Jacobi equations
- Flow and Elastic Networks on the n-Torus: Geometry, Analysis, and Computation
- Automatic differentiation of nonsmooth iterative algorithms
- Constraint Propagation Algorithms for Temporal Reasoning
- Traffic Network Control From Temporal Logic Specifications
- Data Complexity and Rewritability of Ontology-Mediated Queries in Metric Temporal Logic under the Event-Based Semantics