DP Combos: Cross-Domain Combination Principles

• DP Combos are a family of techniques that combine DP formalism with additional mechanisms across domains such as graph theory, optimization, privacy, and model theory.
• In graph theory, DP Combos manifest as DP-coloring, transforming traditional list coloring into a search for transversal independent sets with sharp threshold and counting properties.
• In optimization and differential privacy, DP Combos integrate dynamic programming with constraint propagation or co-design DP-SGD parameters, enhancing computational efficiency and privacy guarantees.

The phrase “DP Combos” appears in several unrelated technical literatures as a shorthand for constructions in which a DP formalism is combined with another mechanism: correspondence constraints in graph coloring, counting and random-cover techniques, constraint propagation in dynamic programming, privacy accounting and utility optimization for DP-SGD, and, more loosely, tameness correspondences in dp-minimal model theory and metaheuristic–deep learning pipelines (Bernshteyn et al., 2016, Marijnissen et al., 17 Mar 2026, Ganesh, 2024, Simon et al., 2019, Assunção et al., 2024). This suggests that the term functions less as a single standardized notion than as a family of domain-specific combination principles.

1. DP-coloring as a graph-theoretic combination framework

In graph theory, DP-coloring, also called correspondence coloring, was introduced by Dvořák and Postle and formalized for graphs and multigraphs by Bernshteyn, Kostochka, and Pron (Bernshteyn et al., 2016). A cover of a graph $G$ is a pair $(L,H)$ in which the sets $L(v)$ are pairwise disjoint, each $H[L(v)]$ is a clique, and for every edge $uv\in E(G)$ the edges between $L(u)$ and $L(v)$ form a matching. An $(L,H)$-coloring is an independent set in $H$ of size $|V(G)|$, so DP-coloring converts coloring into the search for a transversal independent set in an auxiliary graph (Bernshteyn et al., 2016).

This framework strictly generalizes list coloring. Ordinary list coloring is recovered by taking vertices $(L,H)$0 and connecting $(L,H)$1 to $(L,H)$2 whenever $(L,H)$3 and color $(L,H)$4 appears in both lists. Because DP-coloring quantifies over all such matchings rather than only the identity matching, it always satisfies

$(L,H)$5

The data emphasize that the inequalities can be strict even for simple graphs: $(L,H)$6 for every cycle $(L,H)$7, including even cycles, whereas $(L,H)$8 (Bernshteyn et al., 2016).

For multigraphs, the cover condition is modified so that between $(L,H)$9 and $L(v)$0 one has the union of $L(v)$1 matchings. This yields explicit extremal obstructions. The multigraphs $L(v)$2 and $L(v)$3, obtained by replacing each edge of $L(v)$4 or $L(v)$5 by $L(v)$6 parallel edges, are not DP-degree-colorable; in particular, $L(v)$7 has $L(v)$8 (Bernshteyn et al., 2016).

The degree-list theory also has a precise structural classification. If $L(v)$9 is a connected multigraph, then $H[L(v)]$0 is not DP-degree-colorable if and only if each block of $H[L(v)]$1 is one of the graphs $H[L(v)]$2 or $H[L(v)]$3. This is the DP analogue of the Borodin–Erdős–Rubin–Taylor characterization, with “odd cycles” replaced by arbitrary cycle multigraphs (Bernshteyn et al., 2016). The same paper derives DP-critical edge bounds, including

$H[L(v)]$4

for every $H[L(v)]$5-vertex DP-$H[L(v)]$6-critical multigraph $H[L(v)]$7, and for simple DP-$H[L(v)]$8-critical graphs distinct from $H[L(v)]$9,

$uv\in E(G)$0

These results show that one major meaning of “DP combo” is the interaction of correspondence constraints with degree conditions, block structure, and extremal edge counts (Bernshteyn et al., 2016).

2. Counting, deletion–contraction, and graph operations in DP-coloring

A second graph-theoretic use of “DP combos” concerns counting invariants and graph operations. The DP color function

$uv\in E(G)$1

is the minimum number of DP-colorings over all full $uv\in E(G)$2-fold covers, and its dual

$uv\in E(G)$3

is the corresponding maximum over full $uv\in E(G)$4-fold covers (Mudrock, 2021). These quantities satisfy

$uv\in E(G)$5

so they interpolate between worst-case and best-case correspondence structures (Mudrock, 2021).

The deletion–contraction relation for the DP color function is not identical to the chromatic-polynomial recurrence. For a full $uv\in E(G)$6-fold cover $uv\in E(G)$7 of a multigraph $uv\in E(G)$8 and an edge $uv\in E(G)$9,

$L(u)$0

which yields the global inequality

$L(u)$1

Under a non-overlap condition on the matchings for parallel edges, equality holds at the cover level (Mudrock, 2021). This coupling of a minimum-type invariant with a maximum-type invariant is a characteristic DP-coloring phenomenon.

Several exact formulas are known. For a tree $L(u)$2 on $L(u)$3 vertices,

$L(u)$4

For cycles, if $L(u)$5 is even then

$L(u)$6

while if $L(u)$7 is odd then $L(u)$8; the dual values interchange the asymmetry (Mudrock, 2021). These formulas quantify how DP-coloring can reduce or increase counts relative to ordinary coloring depending on the parity of the cycle.

Joins and gluings produce another major cluster of “DP combos.” For the join parameter

$L(u)$9

one has

$L(v)$0

for a graph $L(v)$1 with chromatic number $L(v)$2 and $L(v)$3, and consequently

$L(v)$4

for every $L(v)$5 (Zhang et al., 2021). Thus sufficiently large joins with complete graphs neutralize the additional obstruction carried by arbitrary correspondences.

At the level of the DP color function, joins and vertex-gluings admit exact threshold statements. For cycles,

$L(v)$6

and for vertex-gluings of graphs each of which is either chordal or a cycle,

$L(v)$7

for any vertex-gluing $L(v)$8 (Becker et al., 2021). Cones over disjoint unions of cycles exhibit a finer threshold: $L(v)$9 when there are two $(L,H)$0-cycles in the base and $(L,H)$1 otherwise (Becker et al., 2021). These results show that graph operations can either preserve exact chromatic-polynomial behavior or create new threshold effects specific to DP-covers.

3. Random covers, fractional relaxations, and partial DP-coloring

A probabilistic meaning of “DP combos” arises when DP-coloring is combined with random covers. A random $(L,H)$2-fold DP-cover $(L,H)$3 is constructed by taking $(L,H)$4 and, for each edge $(L,H)$5, choosing an independent uniformly random permutation $(L,H)$6 and matching $(L,H)$7 to $(L,H)$8 (Bernshteyn et al., 2023). This induces a threshold picture governed by the maximum density

$(L,H)$9

The principal asymptotic statement is that random $H$0-fold DP-covers support threshold behavior around $H$1: graphs are non-DP-colorable with high probability when $H$2 is sufficiently smaller than $H$3, and DP-colorable with high probability when $H$4 is sufficiently larger than $H$5, with a sharp threshold for dense enough graphs (Bernshteyn et al., 2023). For graph sequences $H$6 with $H$7, the function

$H$8

is a DP-threshold under mild growth hypotheses, and it is a sharp DP-threshold when $H$9 (Bernshteyn et al., 2023). In particular, for $|V(G)|$0 the sharp threshold is $|V(G)|$1, and for complete multipartite graphs $|V(G)|$2 it is $|V(G)|$3 (Bernshteyn et al., 2023).

For sparser graphs, degeneracy replaces density as the useful control parameter. If $|V(G)|$4 has degeneracy $|V(G)|$5 and

$|V(G)|$6

then $|V(G)|$7 is $|V(G)|$8-DP-colorable from the random cover $|V(G)|$9 with probability at least $(L,H)$00 (Bernshteyn et al., 2023). The proof uses a Greedy Transversal Procedure and a Chernoff–Hoeffding bound for negatively correlated variables. The same paper develops a fractional version, replacing one-element transversals by independent $(L,H)$01-fold transversals in $(L,H)$02-fold covers and defining $(L,H)$03; the fractional threshold picture is cleaner for sparse graphs (Bernshteyn et al., 2023).

Another branch studies partial DP-coloring. For an integer $(L,H)$04, the partial DP $(L,H)$05-chromatic number

$(L,H)$06

is the minimum of $(L,H)$07 over all $(L,H)$08-fold covers $(L,H)$09 of $(L,H)$10 (Kaul et al., 2020). The natural DP analogue of the Partial List Coloring Conjecture asks whether

$(L,H)$11

for all $(L,H)$12. This fails. The cube graph $(L,H)$13 has $(L,H)$14 but

$(L,H)$15

and for each $(L,H)$16 there is a graph $(L,H)$17 on $(L,H)$18 vertices with

$(L,H)$19

and $(L,H)$20 (Kaul et al., 2020). These counterexamples are built via 2-fold covers with twist representations, where the existence of an $(L,H)$21-coloring is controlled by parity conditions on cycles.

At the same time, partial DP-coloring preserves several list-coloring-type inequalities. One has the subadditivity lemma

$(L,H)$22

the general lower bound

$(L,H)$23

and the fact that the linear bound $(L,H)$24 holds for at least half of the values $(L,H)$25 (Kaul et al., 2020). Chordal graphs, series-parallel graphs, and connected subcubic triangle-free graphs other than $(L,H)$26 are partially DP-nice, and for every graph $(L,H)$27 there exists $(L,H)$28 such that $(L,H)$29 is partially DP-nice (Kaul et al., 2020).

4. Dynamic programming combined with constraint propagation

Outside graph coloring, “DP combos” is used for hybrid optimization architectures in which dynamic programming is combined with another model-based paradigm. A central example is Domain-Independent Dynamic Programming with Constraint Propagation, which integrates a general-purpose CP solver into the DIDP framework and evaluates the combination on Single Machine Scheduling with Time Windows, RCPSP, and TSPTW (Marijnissen et al., 17 Mar 2026).

The base DIDP model is a state-transition system with Bellman recursion

$(L,H)$30

DIDP also uses a dominance relation $(L,H)$31 and a dual bound $(L,H)$32 (Marijnissen et al., 17 Mar 2026). The hybrid modifies transition generation: from a DP state $(L,H)$33, it constructs a CP model $(L,H)$34, runs propagation, prunes the state if the CP model is infeasible or if the CP-strengthened bound is already above the best primal bound, and filters successors with a state-specific infeasibility test (Marijnissen et al., 17 Mar 2026). The evaluation shows that constraint propagation significantly reduces the number of state expansions, solving more instances than a DP solver for Single Machine Scheduling and RCPSP, and giving similar improvements for tightly constrained TSPTW instances; runtime results indicate that the benefits of propagation outweigh the overhead for constrained instances, while further reduction of propagation overhead remains a target (Marijnissen et al., 17 Mar 2026).

A closely related case study is the Partial Shop Scheduling Problem, where DP serves as the primary search framework and CP is used as a subroutine for global constraint propagation (Legrand et al., 22 May 2026). The DP state is extended to

$(L,H)$35

where $(L,H)$36 stores dynamic precedence constraints learned by CP (Legrand et al., 22 May 2026). The hybrid transition routine performs the DP earliest-start update, builds a CP model with interval variables and NoOverlap/precedence constraints, runs a Fixpoint procedure, discards infeasible transitions, and adds CP-discovered precedences back into $(L,H)$37 (Legrand et al., 22 May 2026). A CP-powered lower bound is computed by binary search on makespan using repeated propagation, and the original layer-wise DP is replaced by Anytime Column Search. The same DP+CP engine is then reused inside a Large Neighborhood Search scheme in which partial-order schedules are imposed across restarts (Legrand et al., 22 May 2026).

The significance of these results is methodological. In both papers, DP supplies the state-based search backbone, dominance, and caching, while CP supplies global inference, feasibility detection, and lower-bound strengthening (Marijnissen et al., 17 Mar 2026, Legrand et al., 22 May 2026). The combination is presented as domain-independent in one case and as a flexible scheduling architecture in the other. A plausible implication is that “DP combos” in combinatorial optimization now denotes a reusable pattern rather than an isolated heuristic: DP handles structural search, and a second paradigm acts as an inference oracle.

5. Differential privacy composition and proactive DP-SGD design

In the differential privacy literature, “DP combos” refers to how privacy guarantees combine across iterations, sampling schemes, group sizes, and utility objectives. One line of work gives tight group-level $(L,H)$38-DP guarantees for DP-SGD using Mixture of Gaussians mechanisms and privacy loss distributions (Ganesh, 2024).

That analysis models a single Poisson-sampled DP-SGD step for groups of size $(L,H)$39 by a scalar Mixture of Gaussians with sensitivity random variable $(L,H)$40, and a single fixed-batch step by a scalar Mixture of Gaussians with sensitivity $(L,H)$41 (Ganesh, 2024). The vector-to-scalar reduction and PLD composition yield group-level guarantees that are tight up to discretization error, assuming every intermediate iterate is released (Ganesh, 2024). The practical message is that group privacy for DP-SGD should be computed directly at the group level rather than by applying the generic transformation $(L,H)$42, which is both loose and numerically unstable in moderate-$(L,H)$43 regimes (Ganesh, 2024).

A second line, “Proactive DP,” formulates DP-SGD hyperparameter selection as a multiple target optimization problem (Dijk et al., 2021). Instead of reactive accounting after a training plan is fixed, the method starts from a target privacy budget and uses a closed-form guarantee relating DP-SGD parameters to $(L,H)$44. Its central equation is

$(L,H)$45

For $(L,H)$46 and $(L,H)$47, the paper states that DP-SGD is $(L,H)$48-DP if $(L,H)$49 satisfies the above equation together with a lower bound on the number of rounds $(L,H)$50 of order approximately $(L,H)$51, where $(L,H)$52 counts epochs through the total number $(L,H)$53 of gradient computations (Dijk et al., 2021). It also proves a near-tightness statement: if $(L,H)$54 is more than a constant factor approximately $(L,H)$55 smaller than that lower bound, then the $(L,H)$56-DP guarantee is violated (Dijk et al., 2021).

Proactive DP couples this theory to a utility graph and DP calculator. The utility graph estimates the anticipated utility of a candidate clipping/noise pair $(L,H)$57, and the DP calculator then chooses compatible values of $(L,H)$58, batch size, and related parameters under the fixed privacy budget (Dijk et al., 2021). This makes “DP combos” explicit in the design sense: privacy targets, optimization length, batch structure, noise level, and accuracy objectives are tuned jointly rather than sequentially.

6. Other context-dependent uses of the label

In model theory, the phrase appears in connection with dp-minimality and other minimalities. The core result is that a first-order expansion of $(L,H)$59 is dp-minimal if and only if it is o-minimal (Simon et al., 2019). Analogous correspondences are proved for algebraic closures of finite fields, $(L,H)$60-adic fields, ordered abelian groups with only finitely many convex subgroups, and abelian groups equipped with archimedean cyclic group orders (Simon et al., 2019). For ordered abelian groups with finitely many convex subgroups, dp-minimality is characterized by unary definable sets that are finite unions of sets of the form $(L,H)$61, where $(L,H)$62 is convex (Simon et al., 2019). In this setting, “dp combos” denotes tameness correspondences between dp-minimality and classical minimality notions.

A very different usage appears in automated data mining, where a “DP combo” denotes a metaheuristic plus classifier or neural network pipeline (Assunção et al., 2024). The setup searches over label assignments $(L,H)$63 for a small dataset, using a classifier trained on the candidate labels and validation accuracy on a small ground-truth set as the fitness function. The paper instantiates this with GA or SA as the outer optimizer and ANN, SVM, or RF as the evaluator (Assunção et al., 2024). On MNIST, GA with elitism gives some improvement over random labelings, but final accuracies remain around $(L,H)$64–$(L,H)$65, while SA often stays around $(L,H)$66–$(L,H)$67; the central empirical conclusion is that validation accuracy on a small ground-truth dataset is inadequate for correcting labels of other data instances (Assunção et al., 2024). This use is terminologically remote from the graph-theoretic and privacy-theoretic senses, but it preserves the same combination motif: a DP-related outer mechanism plus a second inference component.

Taken together, these literatures show that “DP Combos” is best understood as a polysemous research label. In graph theory it names combinations of DP-covers with structural, counting, probabilistic, and extremal methods (Bernshteyn et al., 2016, Mudrock, 2021, Becker et al., 2021, Bernshteyn et al., 2023, Kaul et al., 2020, Zhang et al., 2021). In combinatorial optimization it denotes dynamic programming enriched by constraint propagation (Marijnissen et al., 17 Mar 2026, Legrand et al., 22 May 2026). In differential privacy it concerns the composition of privacy loss and the co-design of DP-SGD parameters (Ganesh, 2024, Dijk et al., 2021). In model theory and automated data mining, it functions as a looser shorthand for pairings between a dp- or DP-centered formalism and a second structural principle (Simon et al., 2019, Assunção et al., 2024).