Separated-Ordering: A Decompositional Approach

• Separated-Ordering is a framework that decomposes ordering into distinct, isolable components applicable in experimental, algorithmic, and computational settings.
• It enables independent optimization by treating order parameters separately, as seen in DNN accelerators, magnetic sublattices, and forbidden-pattern order problems.
• The approach improves system efficiency by reducing bit transitions in NoC designs and providing clear, actionable steps in order-theoretic optimization.

Separated-ordering is not a single standardized term across research areas. In the cited literature it appears most explicitly as a data-transmission method in NoC-based DNN accelerators, where weights and inputs are ordered according to their own '1'-bit counts, respectively (Chen et al., 30 Aug 2025). In other contexts the phrase is used more descriptively for situations in which ordering is experimentally, combinatorially, or categorically split into distinguishable components: rare-earth and transition-metal magnetic sublattices in multiferroics (Shukla et al., 2013), pairs of constrained linearizations of partial orders (Jiang et al., 2021), or kernel-versus-quotient order in split extensions of preordered groups (Clementino et al., 2022). A plausible synthesis is that separated-ordering denotes frameworks in which order is not treated as a single undifferentiated object, but as a structure with components that can be isolated, compared, or optimized separately.

1. Terminological scope and recurrent structure

One recurrent feature of the literature is that several papers explicitly do not define a formal problem named “Separated-Ordering,” yet supply a closely related formalism. In partial-order alignment, the task is to choose linear extensions of two partial orders on the same ground set so that the resulting total orders maximize common adjacencies or minimize breakpoints (Jiang et al., 2021). In forbidden-pattern graph ordering, the closest explicit analogue is the bipartite side-respecting problem $BIORD(F)$, where the two sides of a fixed bipartition are ordered separately subject to forbidden ordered patterns (Hell et al., 2014). In split extensions of preordered groups, the lexicographic order separates the order on the quotient from the order on the kernel and becomes the central existence criterion for compatible orders (Clementino et al., 2022).

A second recurrent feature is that separation may be physical, algorithmic, or logical. In condensed-matter systems it can refer to different magnetic or structural order parameters with different temperature dependences and different experimental signatures (Shukla et al., 2013). In permutation-pattern and forbidden-ordering theory it can refer to ordinal sums, disjoint sums, or forbidden local interleavings of ordered substructures (Biswas et al., 21 Oct 2025). In optimization and learning it can refer to ordering information acting as a signal distinct from content, as in the decomposition of training dynamics into content and ordering terms (LeDoux, 24 Mar 2026).

This breadth makes the term context-sensitive. In some fields it denotes a named procedure; in others it is best understood as an interpretive label for a decomposition of ordering constraints or order parameters. The common denominator is the explicit treatment of order as something that can be factored into separately meaningful components.

2. Separated ordering in condensed-matter systems

In multiferroic HoFe$_3$(BO$_3$)$_4$, separated-ordering denotes experimentally distinguishable ordering of the Ho and Fe magnetic sublattices (Shukla et al., 2013). Resonant X-ray scattering at the Ho $L_3$ edge isolates the Ho contribution and shows a screw-type order: Ho moments form a basal-plane spiral around the $c$ axis, propagating with a 60$^\circ$ rotation from one crystallographic plane to the next along $c$. High-energy non-resonant X-ray magnetic scattering at $(0\,0\,1.5)$ tracks the basal-plane component dominated by Fe and reveals a distinct Fe spin rotation within the $ab$ plane near $_3$0 K. The temperature sequence is $_3$1–39 K for long-range magnetic order, onset of spontaneous polarization around $_3$2–23 K, and a final spin-reorientation transition at $_3$3–5 K. The Ho resonant intensity scales with $_3$4, while the Fe-dominated non-resonant intensity scales with $_3$5, so different components of the ferroelectric polarization are associated with different magnetic sublattices. The Ho order is initially induced by Fe order, but the rapid growth of the Ho sublattice at lower temperature feeds back on Fe and drives a competing anisotropy-controlled rearrangement. In that setting, “separated” does not mean uncoupled; it means experimentally isolated, temperature-differentiated, and only partially cooperative.

A related but distinct two-stage ordering occurs in kagome spin ice (Chern et al., 2011). There the low-energy manifold contains both microscopic Ising spins and emergent magnetic charges $_3$6. In the short-range model $_3$7, the system displays an intermediate critical phase bounded by Kosterlitz–Thouless transitions. In the dipolar model $_3$8, the intermediate phase instead has long-range staggered charge order, with an upper transition in the 2D Ising class and a lower transition expected asymptotically in the 2D 3-state Potts class. Here ordering is separated because charges order before the spins fully select one of the six $_3$9 magnetic ground states.

In CeFeAsO, the relevant separation is between structural and Fe magnetic transitions (Jesche et al., 2010). The best single crystals show $_3$0 K and $_3$1 K, so $_3$2 K, while earlier reports had values near $_3$3 K. The decrease of $_3$4 with sample quality was used to argue that a large split is not an intrinsic fixed property of the 1111 pnictides. In superconducting phase-separated Cs$_3$5Fe$_3$6Se$_3$7, the main ordered phase and a secondary phase coexist throughout the measured pressure range up to $_3$8 GPa, while the main phase undergoes an $_3$9 Fe-vacancy order-disorder transition near $_4$0–11 GPa with kinetics on the order of hours (Svitlyk et al., 2014). In $_4$1, separated ordering appears either as ferromagnetic domains of about $_4$2 embedded in a G-type antiferromagnetic matrix at $_4$3, or as structural coexistence of orthorhombic Pnma and monoclinic $_4$4 phases carrying G-type and C-type antiferromagnetic orders at $_4$5 (Sharma et al., 2016).

Spatially fragmented variants also occur. In the frustrated honeycomb-lattice $_4$6-$_4$7 Ising antiferromagnet for $_4$8, Monte Carlo simulations indicate frozen stripe-type antiferromagnetic domains separated by zero-energy domain walls, with local order inside domains but no conventional magnetic long-range order across the whole lattice (Žukovič et al., 2019). For driven skyrmions on random pinning, strong pinning and strong Magnus force produce either density phase separation into dense moving bands and depleted regions, or dynamically phase-separated states with nearly uniform density but motion localized in bands (Reichhardt et al., 2018). Across these examples, separated-ordering describes either multiple coupled order parameters with different observables, or ordered regions that remain spatially or dynamically segregated.

3. Separated linearization, pattern avoidance, and ordering complexity

In algorithmic order theory, one of the clearest nearby formalizations is the partial-order alignment problem (Jiang et al., 2021). Given partial orders $_4$9 and $L_3$0 on the same marker set $L_3$1, a linear extension must satisfy

$L_3$2

The optimization problems are Max-Adj, maximizing the number of common adjacencies between chosen linearizations, and Min-Brk, minimizing breakpoints. They are equivalent through

$L_3$3

The paper proves that $L_3$4 and $L_3$5 are APX-hard, and that $L_3$6 and $L_3$7 remain APX-hard even when every bucket of each weak order has size at most two. It also gives a polynomial-time exact algorithm for $L_3$8 and $L_3$9 via dynamic programming over weak-order buckets. In this line of work, separated-ordering is a natural label for choosing two constrained total orders jointly while optimizing agreement on local consecutive structure.

Forbidden-pattern ordering provides another formal analogue (Hell et al., 2014). For a set $c$0 of ordered graph patterns, $c$1 asks whether a graph admits an $c$2-free ordering. In the bipartite version $c$3, a bipartite graph $c$4 with fixed bipartition $c$5 is given, and the task is to order $c$6 and $c$7 separately so that no forbidden bipartite pattern occurs. For $c$8, Theorem 3 states that $c$9 has an $^\circ$0-free ordering of its parts if and only if no strong component of the constraint digraph $^\circ$1 contains a circuit; hence every $^\circ$2 with $^\circ$3 is polynomial-time solvable. When “separated” means side-respecting or partition-respecting, $^\circ$4 is the closest explicit formalism in that paper.

Partially ordered patterns extend this viewpoint through ordinal and disjoint sums of labeled posets (Biswas et al., 21 Oct 2025). The ordinal sum $^\circ$5 preserves the internal orders of $^\circ$6 and $^\circ$7 and forces every element of the first summand below every element of the second. The disjoint sum $^\circ$8 preserves only internal orders and imposes no cross-block comparabilities. The paper reinterprets earlier shape-Wilf-equivalence results as ordinal-sum statements, proves disjoint-sum analogues, establishes $^\circ$9, and completely classifies POPs of sizes $c$0 whose connected components are all chains. The distinction between layered separation ($c$1) and independent separation ($c$2) makes “separated” literal at the level of poset composition.

A broader complexity-theoretic result shows that ordering problems defined by finitely many forbidden ordered subgraphs capture the class $c$3 (Kun et al., 17 Apr 2025). This refutes a general dichotomy conjecture for such ordering problems. At the same time, finite sets of biconnected ordered patterns do satisfy a tractability-versus-$c$4-completeness dichotomy, and a single forbidden biconnected ordered graph is $c$5-complete unless it is the ordered complete graph. A plausible implication is that separated-order constraints built from disconnected or marker-like patterns can be substantially more expressive than local biconnected ordering constraints.

4. Separated-ordering in NoC-based DNN accelerators

The most explicit technical use of the term occurs in NoC-based DNN accelerators (Chen et al., 30 Aug 2025). There, separated-ordering is defined as follows: “weights and inputs are ordered according to their own ‘1’-bit counts, respectively.” Unlike affiliated-ordering, which sorts weights and moves the paired inputs with them, separated-ordering independently sorts the weight stream and the input/activation stream by descending population count. Because the positional pairing between inputs and weights is broken, a minimal-bit-width index is required for recovery.

The method operates at flit granularity. A flit is divided into an input half and a weight half, and separated-ordering applies the same descending '1'-bit-count principle independently to each half. The mathematical rationale is derived from the expected bit-transition count between consecutive 32-bit words and then between consecutive flits. For two flits containing $c$6 words each, with popcounts $c$7 and $c$8,

$c$9

Since the sums of one-bit counts are fixed for a given payload, minimizing $(0\,0\,1.5)$0 is equivalent to maximizing

$(0\,0\,1.5)$1

The paper provides the proof for the generic count-based ordering principle and then applies that principle independently to the input and weight streams. It does not provide a separate theorem proving global optimality of separated-ordering for the full DNN scheduling problem.

Architecturally, the ordering unit is placed near off-chip memory rather than inside routers, and ordering latency is intended to be hidden between layers (Chen et al., 30 Aug 2025). The same hardware used for affiliated-ordering can be used for separated-ordering “with double time consumption.” In full NoC experiments, the methods are compared as O0 (baseline), O1 (affiliated-ordering), and O2 (separated-ordering). Across different NoC sizes, separated-ordering yields 23.30% to 32.01% bit-transition reduction for float-32 data and 16.95% to 35.93% for fixed-8 data. Across DNN models including LeNet and a DarkNet-like model, the strongest reductions are up to 35.93% for LeNet and 40.85% for DarkNet. The headline NoC results are up to 32.01% BT reduction for float-32 and 40.85% for fixed-8. Using a synthesized link model with bit-transition energy $(0\,0\,1.5)$2, the paper estimates that with 40.85% BT reduction link power falls from 155.008 mW to 91.688 mW, or from 476.672 mW to 281.951 mW in Banerjee’s model. In this literature, separated-ordering is therefore a concrete low-level transmission policy: independent ordering of semantically coupled streams to reduce switching activity.

5. Abstract order-theoretic and topological uses

In the theory of split extensions of preordered groups, the central separated-order construction is the lexicographic order on a semidirect product (Clementino et al., 2022). For a split extension

$(0\,0\,1.5)$3

the product cone is $(0\,0\,1.5)$4, while the lexicographic cone is

$(0\,0\,1.5)$5

Proposition 3.1 states that a positive cone $(0\,0\,1.5)$6 is compatible with the split extension if and only if

$(0\,0\,1.5)$7

Theorem 3.2 then shows that a compatible order exists if and only if the lexicographic order is compatible, which is equivalent to every action map $(0\,0\,1.5)$8 being monotone and every $(0\,0\,1.5)$9 acting pointwise as $ab$0. The lexicographic order is thus the maximal compatible separated order, while the submonoid generated by $ab$1 is the minimal compatible one. The same paper proves that the Split Short Five Lemma holds for stably strong split extensions.

A topological use of separation language appears in left-separated spaces (Soukup et al., 2016). A well-order $ab$2 of a space $ab$3 is left-separating if $ab$4 is closed for every $ab$5, and $ab$6 is the minimum order type of such a well-order. For a regular cardinal $ab$7 and each $ab$8, there is a $ab$9 space $_3$00 with

$_3$01

Under stronger hypotheses there are 0-dimensional examples, and in additional cases there are locally compact, locally countable, 0-dimensional examples with the same exact type. The union of two left-separated spaces is not necessarily left-separated, but if $_3$02 is countably tight, $_3$03, and $_3$04, then $_3$05 is left-separated and

$_3$06

It is also consistent that a first countable, 0-dimensional space is not left-separated in the ground model but becomes left-separated in type $_3$07 after c.c.c. forcing.

Two neighboring but non-identical notions further illustrate the breadth of separation vocabulary. In group theory, a group $_3$08 is order separable if non-conjugate and non-inverse-conjugate elements can be sent to a finite quotient with different orders; for free products, $_3$09 is order separable if and only if $_3$10 and $_3$11 are order separable (Yedynak, 2010). In linear-order theory, there exist non-isomorphic linear orders $_3$12 and $_3$13 with

$_3$14

for $_3$15, so each order divides the other on both left and right under lexicographic product (Ervin, 2018). These are not instances of the NoC or multiferroic notion, but they show how separation language repeatedly arises when order is decomposed into asymmetric sides, factors, or stages.

6. Ordering as an independent information channel

A recent learning-theoretic development makes the separation of order from content completely explicit (LeDoux, 24 Mar 2026). In modular addition with $_3$16, the model, initialization, optimizer, hyperparameters, training set, epoch budget, and compute budget are fixed; only the permutation of the $_3$17 training pairs changes. The full input space has $_3$18 pairs, so the training set is about 0.3% of the input space, and test performance is measured on 1,000,000 held-out pairs. Under Stride ordering, test accuracy reaches 99.5% by epoch 487; under Fixed-Random, it reaches 99.5% by epoch 659; the IID Random baseline reaches only 0.30% after 5,000 epochs; and the adversarial Target ordering remains at 0.01% test accuracy with 0.02% train accuracy. The paper interprets this as evidence that order is a distinct information channel.

The formal mechanism is a decomposition of the gradient seen by batch $_3$19 after an update on batch $_3$20: $_3$21 The first term is the content term; the second is the ordering or entanglement term. A counterfactual shuffled baseline is then used to define

$_3$22

Measured this way, the ordering fraction accounts for 83% of gradient energy under Stride, 84% under Fixed-Random, 87% under Random, and 89% under Target. In the Stride case with $_3$23, the learned embedding spectrum concentrates on a fundamental Fourier mode

$_3$24

and the same fundamental emerges across all tested seeds. The paper’s phrase “the order is the message” is therefore literal: ordering information is separable from example content and can determine both optimization trajectory and learned representation. A plausible implication is that separated-ordering can also mean the explicit isolation of sequence structure as its own causal variable.

Taken together, these literatures show that separated-ordering is best understood not as one universally fixed definition, but as a family of constructions in which order is decomposed into experimentally isolated sublattices, independently sorted data streams, separately constrained linearizations, lexicographically prioritized components, or content-independent temporal signals. The persistence of the idea across such different settings suggests that “separated” names a structural operation on order itself: isolating the parts of an ordering problem that can be treated as distinct without becoming irrelevant.