Recirculate-and-Truncate: Multi-Domain Technique

• Recirculate-and-Truncate is a technique that uses repeated re-injection combined with systematic truncation to emulate byte-wise payload scanning and manage state efficiently.
• It spans multiple domains—from programmable data planes for DPI to spatial microsimulation and multi-objective optimization—demonstrating broad applicability.
• The method balances repeated processing with truncation to control memory growth and stabilize dynamics under hardware and analytic computational constraints.

Recirculate-and-truncate denotes, in its clearest and explicitly named form, a multi-pass workaround for programmable data planes in which a packet is repeatedly re-injected into the ingress pipeline and shortened on each pass, allowing a single-pass parser to emulate byte-wise payload scanning for deep packet inspection (DPI) and next-generation firewalling (Singh et al., 29 Jul 2025). In broader technical exposition, the same label has been used descriptively for procedures that combine repeated traversal, restart, or reuse of an object with deliberate truncation, thresholding, or fading-memory attenuation in spatial microsimulation, multi-objective optimisation, Monte Carlo reinforcement learning, stochastic routing, cosmological amplitude calculation, laser–plasma acceleration, and multimode-fiber recurrent computation (Lovelace et al., 2013, Cui et al., 2 Apr 2025, Poiani et al., 2023, He et al., 2024, Jiang et al., 2019, DuBois et al., 2017, Eşlik et al., 22 Feb 2026).

1. Terminology and scope

In the cited literature, the phrase is explicitly used as a named architectural workaround in programmable-switch security. In several other domains, it appears as an expository mapping rather than as the paper’s own formal method name. This distinction is important because the shared intuition is structural rather than terminological: repeated processing creates multiple opportunities to accumulate, inspect, or transform state, while truncation bounds that process by discarding a prefix, clipping a variable, limiting an archive, or attenuating historical influence.

Domain	Recirculation object	Truncation mechanism
Programmable switches	Packet re-injected through pipeline	First payload byte removed per pass
Spatial microsimulation	Individuals with integer replications	Fractional weights sampled after truncation
MOEAs	Archive updated after each arrival	Immediate size reduction to $\mu$
Stochastic $k$-TSP	Repeated orienteering tours	Rewards or costs thresholded by scale
Monte Carlo RL	Episodes restarted under a DCS	Trajectories collected at bounded horizons
Optical or physical systems	Feedback loop or returning carriers	Natural fading, sheath depletion, or $q \to 0$ truncation

Taken together, these usages suggest a general design pattern: a state or object is revisited under controlled repetition, but each revisit is paired with a rule that suppresses uncontrolled growth, long-range dependence, or incompatible state accumulation.

2. The canonical meaning in programmable data planes

In programmable data planes, recirculate-and-truncate is a response to the architectural constraints of P4 targets. A P4 parser is single-pass and cannot loop over arbitrary payload bytes; it “essentially fetches bit-slices of a given length starting from a given offset,” which is adequate for headers but not for application-layer payload inspection. The workaround is to send the packet back through the pipeline multiple times and, at each pass, emit a clone whose payload is shortened by one byte. On pass $i$, the parser therefore sees payload$[i..]$, which realizes a sliding-window traversal of the payload at line rate. Across passes, a deterministic finite automaton (DFA) stored in ternary match tables maps the current state and input character to the next state, so pattern matching and regex-style DPI become feasible despite fixed pipelines and restricted instruction sets (Singh et al., 29 Jul 2025).

The survey identifies P4’s clone and recirculate primitives as the relevant mechanisms. DeeP4R is described as implementing DPI by recirculating a cloned copy of the packet through the switch pipeline while leaving the original untouched; if the clone matches no known patterns, the original is forwarded, otherwise it is discarded. PPS is cited as an earlier DFA-based string-search system on PISA, achieving “line-rate processing at 10–100 Gbps” and supporting “thousands of keyword patterns.” BOLT addresses DFA state explosion through two optimizations: $k$-stride, which consumes multiple input bytes per DFA step, and bit-slicing, which compacts transition logic. The same survey distinguishes this combined technique from recirculation alone: AES via “scrambled lookup tables” uses packet recirculation to simulate multiple cryptographic rounds and reports throughput of 10.92, 8.76, and 7.37 Gbps for AES-128, AES-192, and AES-256, respectively, but does not require truncation because the goal is not byte-wise payload scanning.

The importance of the technique follows directly from hardware limits. The survey places it on PISA/PSA targets with few hundred MB SRAM and a few MB TCAM, and notes that early RMT designs dedicated over 50% of chip area to memory. Recirculate-and-truncate therefore belongs to a larger class of lookup-table workarounds: each pass reduces a forbidden computation to a legal table lookup. Its principal liabilities are also table-centric, especially DFA state explosion, stage pressure, and the memory cost of representing large pattern sets.

3. Repeated selection under truncated surrogates

A closely related logic appears in spatial microsimulation, although the 2013 paper does not explicitly name a recirculate-and-truncate method. Its proposed TRS method—“truncate, replicate, sample”—integerises iterative proportional fitting weights by separating each weight $w_i$ into an integer replication component $r_i=\lfloor w_i \rfloor$ and a fractional component $f_i=w_i-\lfloor w_i \rfloor$. Individuals are first replicated exactly $r_i$ times, producing $k$0 deterministic clones, and the remaining population deficit $k$1 is filled by sampling without replacement with probabilities $k$2. The method was reported as more accurate than simple rounding, threshold, counter-weight, and proportional-probability alternatives: Pearson’s $k$3 was 0.9992 for TRS versus 0.9989, 0.9978, 0.9960, and 0.9911; total absolute error was 28,862; SAE was 3.4%; 39.2% of simulated values were more than 5% from actual; and $k$4 marked 5.5% of cells significant at 5%. Counter-weight, proportional probabilities, and TRS achieved exact zonal population totals in the reported scenario (Lovelace et al., 2013).

In multi-objective optimisation, the phrase has likewise been used descriptively for immediate archive truncation. The 2025 study compares three timings for keeping an archive of nondominated solutions of capacity $k$5: immediate truncation after each new solution, batch truncation, and deferred truncation from an unbounded archive after all 5,000 points have been seen. Over 31 random shuffles of 3-objective simplex and inverted-simplex fronts, immediate truncation tended to be best and unbounded truncation often worst. For NSGA-II on the simplex front, IGD was approximately 0.0466 for immediate truncation, 0.0480 for batch truncation, and 0.2229 for deferred truncation; on the inverted simplex, deferred truncation was about 0.2370. SMS-EMOA showed smaller but consistent differences, with immediate truncation still best. MOEA/D was insensitive to timing, with identical IGD values across strategies, approximately 0.0372 on the simplex and 0.0621 on the inverted simplex. The paper’s interpretation is that one-by-one removal minimizes greedy error accumulation, whereas deferred truncation magnifies it (Cui et al., 2 Apr 2025).

Stochastic $k$6-TSP provides a third instance. There the named ideas are “repetitions” and “critical scaling,” but their combination matches the same structural pattern: truncate stochastic rewards or costs at logarithmically many scales, run a constant number of repeated orienteering tours at each scale, and keep only tours near a critical threshold. For Stoch-Reward $k$7-TSP, the scale-$k$8 surrogate is $k$9; for Stoch-Cost $q \to 0$0-TSP it is $q \to 0$1. The algorithms use $q \to 0$2, $q \to 0$3 repetitions, a “rich” threshold $q \to 0$4, and retain only the tours at $q \to 0$5 and $q \to 0$6. The paper proves non-adaptive $q \to 0$7-approximation algorithms for both problems and hence $q \to 0$8 adaptivity gaps, using Freedman’s and Jogdeo–Samuels’ inequalities to justify truncated expectations as faithful surrogates (Jiang et al., 2019).

4. Temporal recurrence, restart, and fading memory

In Monte Carlo reinforcement learning, a priori truncation of trajectories can be understood as recirculating interaction budget toward earlier timesteps, where discounting gives rewards larger effective weight. The paper formalizes a data collection strategy $q \to 0$9 satisfying $i$0, and the associated per-timestep sample counts $i$1. Its on-policy estimator remains unbiased when rewards at timestep $i$2 are divided by $i$3, and its confidence width is bounded by

$i$4

with

$i$5

Because $i$6 decreases with $i$7, the approximately optimal allocation satisfies $i$8. This produces many short rollouts and fewer long ones, and the same structure is used to extend POIS into TT-POIS. Experiments on Dam Control, Reacher, and a Multi-Echelon Supply Chain, with budgets 8,640, 8,000, and 3,900 per iteration and discount factors $i$9 and $[i..]$0, showed consistent improvements over POIS, with larger gains at smaller $[i..]$1 (Poiani et al., 2023).

A physical analogue appears in multimode-fiber recurrent computation. Successive frames are encoded on three delayed optical channels, combined in a 200/220 $[i..]$2m multimode fiber supporting on the order of $[i..]$3 modes, and fed into a passive 50/50 loop. The loop length is 5 m, the round-trip delay is approximately 24 ns, and the coupler imposes a per-pass amplitude factor of $[i..]$4, i.e. power $[i..]$5 even before additional losses and dispersion. The unrolled recursion shows that the two previous frames enter with power weights proportional to $[i..]$6 and $[i..]$7, so recurrence is implemented by passive recirculation and truncation by exponentially decaying memory. No optical element is trained, and there is no electronic feedback. With a linear ridge-regression readout on measured speckle intensities, the system achieved one-step MSE 0.084 on the Santa Fe laser time series, 98.33% accuracy on KTH human action recognition, and 94.74% when scene was fixed and only action varied; steering-angle prediction and JIGSAWS surgical-skill assessment were also demonstrated qualitatively (Eşlik et al., 22 Feb 2026).

5. Analytic and plasma-physical realizations

In cosmological amplitude calculations, recirculation and truncation appear as a two-step analytic strategy rather than as a data-processing heuristic. The method begins by decomposing a graph into basic time integrals indexed by directed graphs. For a tree with $[i..]$8 edges, the decomposition yields exactly $[i..]$9 basic time integrals; loop graphs have fewer because some orientations are incompatible. Differentiation with respect to vertex energies produces first-order edge-contraction equations in which each derivative replaces an edge by a contracted graph, and the resulting hierarchy can be assembled into a complete differential system. For tree graphs the system can be transformed into canonical differential equations of size $k$0. The “recirculate” step then solves the hierarchy recursively by expressing an $k$1-site integral as a one-fold integral over an $k$2-site contracted integral, iterating to Euler–Mellin representations and generalized hypergeometric functions. The “truncate” step is the de Sitter limit $k$3, where the differential equations truncate analogously to canonical differential equations for Feynman integrals at integer dimensions, yielding complete multiple-polylogarithm symbols. For one-loop $k$4-gons, the basis size is reported as $k$5 (He et al., 2024).

A more literal physical recirculation-plus-truncation mechanism is identified in target normal sheath acceleration. One-dimensional particle-in-cell simulations show that fast electrons generated in low-contrast, few-10-fs laser–foil interactions recirculate through the target and modulate the rear-side charge and current densities. Because the electron source lasts only for the laser pulse duration, periods arise in which most recirculating electrons are in transit away from the target rear; the rear sheath is then transiently truncated. In a 12 $k$6m target, the rear sheath field drops from about 5 TV/m to about 0.2 TV/m for roughly 30 laser periods, producing an early spectral peak that later spreads into a 3–5 MeV plateau. The controlling scale is the recirculation period

$k$7

with $k$8 for relativistic electrons. The reported cases were $k$9 fs for a 12 $w_i$0m target with $w_i$1 fs, which produced a clear plateau; $w_i$2 fs for a 3 $w_i$3m target with the same pulse length, which produced only weak modulation and no plateau; and a 12 $w_i$4m target with $w_i$5 fs, where no plateau formed because the sheath was not depleted (DuBois et al., 2017).

6. Common structure, limitations, and open problems

Across these literatures, recirculate-and-truncate does not denote a single algorithmic primitive. In programmable switches it is an explicit workaround for single-pass parsing; in microsimulation it separates integer replication from fractional sampling; in MOEAs it means immediate archive control; in stochastic routing it combines thresholding with repeated tours; in RL it redistributes budget across truncated horizons; in cosmological amplitudes it is a recursive differential-and-limit procedure; and in optical or plasma systems it can arise directly from feedback and attenuation or from carrier recirculation and sheath depletion. A common misconception is therefore to treat truncation as mere data loss. In these papers, truncation often serves the opposite role: it isolates a meaningful subproblem, preserves exact totals, preserves unbiasedness, or stabilizes a recurrent process.

Another misconception is to equate recirculation alone with the combined method. The programmable-switch survey is explicit that recirculation by itself can implement multi-round AES, whereas recirculate-and-truncate is specifically what makes payload scanning possible because truncation advances the inspection window deterministically (Singh et al., 29 Jul 2025). A related optimisation misconception is that more information always helps if truncation is postponed: the MOEA study shows that an unbounded archive followed by late truncation is often worse than immediate truncation, because one-by-one greedy errors accumulate at scale (Cui et al., 2 Apr 2025).

The main limits are domain-specific. Programmable-switch DPI remains bounded by DFA state explosion, stage budgets, and lookup-table memory; the same survey identifies open directions in complete in-data-plane cryptographic endpoints, payload-aware ML security, and distributed memory/computation. In cosmological amplitudes, the current framework treats conformal scalars with non-derivative, time-dependent self-interactions in power-law FRW backgrounds, and extending it to spins, derivative couplings, and fully integrated loop amplitudes remains open (He et al., 2024). In the multimode-fiber reservoir, memory depth is physically finite because the per-pass gain is strictly below unity and additional loss and dispersion shorten the effective horizon; the reported energy-efficiency advantage is qualitative rather than numerically quantified (Eşlik et al., 22 Feb 2026). In TNSA, later recirculation-driven modulations are expected to be diminished in realistic multidimensional geometry by transverse electron losses (DuBois et al., 2017).

Taken as a family of techniques rather than a single formalism, recirculate-and-truncate names a recurring engineering and analytic motif: repeated access to a state is made useful only when paired with a truncation rule that controls representation size, stabilizes dynamics, or converts an otherwise intractable operation into one that the underlying architecture can actually support.