Minimal Witnessing Subsystems

• Minimal witnessing subsystems are domain-dependent constructs that certify properties of a larger system via a reduced subobject preserving essential semantic content.
• They employ varied criteria—such as support size, invariant strength, and state count—to achieve minimality in domains like quantum spectral factorization, probabilistic verification, and database query preservation.
• The methodology has practical implications including NP-hardness in certain scenarios and the development of algebraic and optimization techniques to isolate critical system structures.

Searching arXiv for the cited papers to ground the article and verify metadata. {"query":"id:(Loizeau et al., 2024) OR id:(Vieira et al., 2023) OR id:(Funke et al., 2019) OR id:(Jantsch et al., 2021) OR id:(Jantsch et al., 2020) OR id:(Jantsch et al., 2020) OR id:(Seibt, 6 Jul 2025)", "max_results": 10} “Minimal witnessing subsystems” denotes a family of domain-dependent constructions in which a smaller subobject certifies a property of a larger system while preserving the semantic content relevant to that property. In the literature, the witness may be a sub-database that preserves all query answers, a sub-MDP or sub-PTA that still satisfies a reachability threshold, an autocatalytic core inside a chemical reaction network, a bounded set of points certifying a tangle, or a minimal dynamical subsystem realizing prescribed complexity or mean dimension. In quantum settings, the same role is played by global data that certifies hidden compositional structure, such as spectral factorization or temporal-correlation bounds on effective environment dimension (Loizeau et al., 2024, Vieira et al., 2023, Funke et al., 2019).

1. Domain-dependent meanings of minimal witnesses

The common pattern is reduction under semantic preservation: a witness is smaller than the ambient object, but still suffices to establish the target statement. What counts as “smaller,” however, is highly context-sensitive. Some literatures minimize support size, some minimize states or locations, some minimize invariant strength or volume, and some use inclusion-minimality rather than minimum cardinality.

Domain	Ambient object	Witness object
Quantum mereology	Global Hamiltonian spectrum	Outer-sum or DOS factorization
Temporal open quantum systems	Probe-only multi-time statistics	Minimum effective environment dimension
Probabilistic verification	DTMC, MDP, PTA	Witnessing subsystem
Databases	Relational database	Sub-database $D' \subseteq D$
Chemical reaction networks	Stoichiometric matrix / König graph	Autocatalytic core
Tangles	Oriented separation system	Witnessing set $W \subseteq V$
Symbolic/topological dynamics	Subshift or Bernstein space	Minimal subsystem

In probabilistic verification, a witness is explicitly a subsystem that still satisfies a lower-bounded reachability property; in databases, a witness is a sub-database satisfying $Q(D') = Q(D)$; in autocatalysis, the closest exact analogue is an autocatalytic core, namely an autocatalytic matrix with no proper autocatalytic submatrix; and in tangles, a witnessing set $W\subseteq V$ certifies the 3-wise consistency condition by hitting every triple intersection required by the tangle axioms (Funke et al., 2019, Hu et al., 2023, Golnik et al., 24 Nov 2025, Seibt, 6 Jul 2025).

A central distinction across these papers is between existential witnesses and reconstructive witnesses. Some witnesses merely certify that some hidden structure must exist. Others return an explicit reduced object. This suggests that “minimal witnessing subsystem” is best read as an umbrella notion rather than a single formal definition.

2. Spectral and temporal quantum witnesses

In “quantum mereology,” subsystem structure is posed as a spectral factorization problem. For a Hilbert space bipartition, an exact noninteracting decomposition exists when there are $U$, $H_A$, and $H_B$ such that

$UHU^\dagger = H_A \otimes 1_B + 1_A \otimes H_B.$

The paper proves the exact equivalence

$$UHU^\dagger = H_A\otimes 1_B + 1_A\otimes H_B \Longleftrightarrow E_{ij}=A_i+B_j \Longleftrightarrow Z(t)=Z_A(t)Z_B(t) \Longleftrightarrow \rho(E)=(\rho_A * \rho_B)(E),$$

so decomposing a system into subsystems is equivalent to decomposing its spectrum into other spectra (Loizeau et al., 2024). In this setting, the minimal exact witness of nontrivial subsystem structure is therefore an outer-sum factorization of the spectrum, equivalently a product factorization of the generating function or a convolution factorization of the density of states.

The same paper argues that in local many-body systems the leading Gaussian density of states is not a minimal witness of a finite elementary decomposition. Using moment expansions for commuting or asymptotically commuting local terms, it obtains Gaussian moments at leading order and identifies finite-size non-Gaussian corrections as the informative part. The normalized fourth moment obeys

$\mu_4^N \approx 3-\frac{2}{M},$

and for geometrically local models $W \subseteq V$0, so

$W \subseteq V$1

The strongest claim is therefore not about reconstructing a tensor-product structure uniquely, but about witnessing the number of elementary local constituents from the first non-Gaussian correction (Loizeau et al., 2024).

A different quantum notion of minimal hidden subsystem appears in open-system dynamics. There the observed object is a multi-time sequence distribution

$W \subseteq V$2

and the witness asks for the smallest effective environment dimension compatible with the data. For a fixed sequence $W \subseteq V$3, the bound

$W \subseteq V$4

is computed through a convergent SDP hierarchy over symmetric extensions. If an experiment yields $W \subseteq V$5, then no model with environment dimension $W \subseteq V$6 explains the data (Vieira et al., 2023).

The paper’s explicit qubit example is a direct dimension witness. For $W \subseteq V$7, $W \subseteq V$8, and sequence $W \subseteq V$9, the hierarchy gives

$Q(D') = Q(D)$0

at level $Q(D') = Q(D)$1, while a qutrit environment achieves

$Q(D') = Q(D)$2

Hence any observed value $Q(D') = Q(D)$3 certifies $Q(D') = Q(D)$4 (Vieira et al., 2023). Here the minimal witnessing subsystem is not reconstructed directly; what is witnessed is the minimum effective hidden memory dimension.

3. Probabilistic verification and timed automata

In finite probabilistic verification, a witnessing subsystem is a smaller DTMC or MDP that still satisfies a reachability threshold. For lower bounds on $Q(D') = Q(D)$5, the key algebraic objects are Farkas certificate polytopes: $Q(D') = Q(D)$6

$Q(D') = Q(D)$7

The central theorem is that witnesses are exactly supports of points, and equivalently supports of vertices, in the appropriate certificate polytope. A minimal witness is obtained precisely from a vertex with a maximal number of zeros (Funke et al., 2019). This turns minimal witnessing-subsystem search into a sparsity problem over a polytope.

That geometric reformulation does not make the problem easy. The decision version is NP-complete even for acyclic DTMCs, and the low-tree-width line of work strengthens the negative result: the witness problem remains NP-hard for Markov chains with $Q(D') = Q(D)$8, and hence also for $Q(D') = Q(D)$9, despite the existence of polynomial-time algorithms when the underlying graph is a tree (Jantsch et al., 2021). The same paper gives an exact dynamic-programming style algorithm for a given directed tree partition. It summarizes child subtrees by interface-value vectors, composes them bottom-up, and removes dominated partial subsystems via convex-hull reasoning.

The tool-oriented counterpart is SWITSS, which implements exact MILP and iterative LP methods for DTMCs and MDPs. The exact route adds binary variables to the certificate LP and returns genuinely minimal witnesses under the selected objective; the heuristic route uses quotient-sum reweighting to obtain small witnesses quickly. SWITSS supports state-based minimization and label-based minimization, can render returned subsystems graphically, and returns a certificate that can be independently checked by verifying the relevant Farkas inequalities (Jantsch et al., 2020).

For probabilistic timed automata, the witness notion must account for uncountable timed state spaces. The paper therefore distinguishes weak and strong subsystems: weak subsystems suffice for $W\subseteq V$0, while strong subsystems are required for $W\subseteq V$1. It introduces three notions of minimality,

$W\subseteq V$2

capturing number of locations, invariant strength, and total invariant volume, respectively (Jantsch et al., 2020). The technical bridge is a PTAB quotient together with Farkas certificates on the quotient MDP. To lift certificate supports back to timed subsystems, the paper introduces the DBM operation

$W\subseteq V$3

whose valuation set is the smallest zone containing the union. This yields exact single-exponential procedures for loc-minimal, inv-minimal, and vol-minimal witnesses relative to the quotient size, while volume comparison is substantially harder: deciding $W\subseteq V$4 is PP-hard (Jantsch et al., 2020).

4. Database, chemical, and separation-based witnesses

In databases, the smallest witness problem asks for a sub-database $W\subseteq V$5 such that

$W\subseteq V$6

and $W\subseteq V$7 is minimum. For conjunctive queries without self-joins, the exact complexity admits a dichotomy: SWP is polynomial-time solvable if and only if the query has head-cluster property, unless $W\subseteq V$8. Approximation has a second dichotomy: constant-factor approximation is possible if and only if the query has head-domination property, while queries without head-domination are not approximable within a $W\subseteq V$9 factor unless $U$0 (Hu et al., 2023). In this literature, the minimal witnessing subsystem is literal tuple support preserving full query semantics.

For chemical reaction networks, the closest notion is the autocatalytic core. The paper defines an autocatalytic matrix as one admitting $U$1 with $U$2, together with the sign condition that each column has both a negative and a positive entry. An autocatalytic core is an autocatalytic matrix with no proper autocatalytic submatrix (Golnik et al., 24 Nov 2025). This is explicitly an inclusion-minimal notion, not a globally minimum-cardinality one. The paper further shows that irreducible autocatalytic Metzler CS-matrices can be characterized graph-theoretically by “fluffles,” and it develops algorithms to enumerate irreducible autocatalytic subnetworks and, more efficiently, autocatalytic cores. In the modified E. coli core network, the complete analysis found $U$3 autocatalytic classes among $U$4 CS-equivalence classes, $U$5 autocatalytic Metzler matrices, and $U$6 autocatalytic cores (Golnik et al., 24 Nov 2025).

For tangles, the witness object is not a subfamily of separations but a bounded set of points. A set $U$7 witnesses a tangle $U$8 if

$U$9

The main theorem proves that every $H_A$0-tangle has a witnessing set of size at most

$H_A$1

improving an earlier result of Grohe and Schweizer; conversely, for every $H_A$2, there exists a $H_A$3-tangle whose smallest witnessing set has size

$H_A$4

The proofs isolate the minimal oriented separations of the tangle as the structurally relevant data, but they do not establish a smallest separation subsystem in the strong sense (Seibt, 6 Jul 2025).

These three literatures illustrate three different meanings of “minimal”: minimum tuple support, inclusion-minimal stoichiometric submatrix, and minimum-cardinality set of witness points. A plausible implication is that the phrase “minimal witnessing subsystem” is intrinsically representation-dependent.

5. Minimal subsystems in dynamical systems

In symbolic dynamics, the phrase itself is not standard, but the closest formal object is the minimal subsystem of a transitive subshift. The paper on transitive subshifts with linear complexity bounds how many distinct minimal subsystems can occur as a function of the complexity $H_A$5. If $H_A$6 is the orbit closure of a recurrent point and has $H_A$7 proper minimal subsystems, exactly $H_A$8 of which are infinite, then

$H_A$9

$H_B$0

For general transitive systems with $H_B$1,

$H_B$2

The same paper also proves that if $H_B$3 is recurrent, transitive, and aperiodic with

$H_B$4

then $H_B$5 is uniquely ergodic (Dykstra et al., 2019). Here minimal subsystems behave as witnesses of persistent invariant structure, and low complexity constrains how many such witnesses can coexist.

A constructive variant appears in Bernstein spaces. For the shift on $H_B$6, the ambient mean dimension is $H_B$7, and for every

$H_B$8

there exists a minimal subsystem $H_B$9 with

$UHU^\dagger = H_A \otimes 1_B + 1_A \otimes H_B.$0

Likewise, for the real-valued space $UHU^\dagger = H_A \otimes 1_B + 1_A \otimes H_B.$1, every

$UHU^\dagger = H_A \otimes 1_B + 1_A \otimes H_B.$2

is realized by a minimal subsystem of that mean dimension (Zhao, 2022). The construction passes through a symbolic skew-product $UHU^\dagger = H_A \otimes 1_B + 1_A \otimes H_B.$3, a controlled density of free coordinates, and an interpolation map into a band-limited function space. In this setting, a minimal subsystem is not a compressed certificate of an external property; it is itself the object realizing the prescribed invariant.

Taken together, these papers show two distinct dynamical uses of minimal subsystems: as objects whose multiplicity is constrained by complexity, and as constructive realizers of prescribed mean dimension.

6. Common themes, ambiguities, and limits

Across the surveyed work, several recurrent themes appear. First, exact witnesses are often algebraic. Examples include

$UHU^\dagger = H_A \otimes 1_B + 1_A \otimes H_B.$4

for quantum spectral factorization; $UHU^\dagger = H_A \otimes 1_B + 1_A \otimes H_B.$5 and $UHU^\dagger = H_A \otimes 1_B + 1_A \otimes H_B.$6 for probabilistic certificates; and $UHU^\dagger = H_A \otimes 1_B + 1_A \otimes H_B.$7 for database witnesses (Loizeau et al., 2024, Funke et al., 2019, Hu et al., 2023). Where exact reconstruction is unavailable or computationally hard, the literature turns to approximate or asymptotic witnesses: SDP upper bounds on temporal correlations, constant-factor or logarithmic approximations for SWP, and LP/MILP heuristics for small probabilistic subsystems (Vieira et al., 2023, Hu et al., 2023, Jantsch et al., 2020).

Second, minimality is not uniform. In finite probabilistic systems it is usually cardinality of states or support; in PTAs it may mean minimum locations, strongest invariants, or least invariant volume; in autocatalysis it means no proper autocatalytic submatrix; in tangles it means minimum-cardinality witnessing set; and in databases it means minimum number of tuples (Jantsch et al., 2020, Golnik et al., 24 Nov 2025, Seibt, 6 Jul 2025, Hu et al., 2023). Confusing these notions leads to real mathematical differences: the autocatalysis paper explicitly distinguishes autocatalytic cores from “minimal autocatalytic subnetworks” that are only reaction-minimal, and the PTA paper shows that $UHU^\dagger = H_A \otimes 1_B + 1_A \otimes H_B.$8, $UHU^\dagger = H_A \otimes 1_B + 1_A \otimes H_B.$9, and $$UHU^\dagger = H_A\otimes 1_B + 1_A\otimes H_B \Longleftrightarrow E_{ij}=A_i+B_j \Longleftrightarrow Z(t)=Z_A(t)Z_B(t) \Longleftrightarrow \rho(E)=(\rho_A * \rho_B)(E),$$0 are not equivalent (Golnik et al., 24 Nov 2025, Jantsch et al., 2020).

Third, non-uniqueness is often intrinsic. In quantum mereology, an exact Gaussian density of states can be partitioned into “an infinite number of infinitely small non-interacting subsystems,” and even GOE spectra can be approximately partitioned with exponentially small spectral error, so spectral data alone are often existential rather than reconstructive (Loizeau et al., 2024). In probabilistic verification, there may be several distinct minimum-cardinality witnesses of the same optimal size, and tool-supported objectives such as label-minimality can select qualitatively different witnesses from state-minimality (Jantsch et al., 2021, Jantsch et al., 2020). In tangles, guiding sets and witnessing sets are complementary rather than equivalent, and a tangle may have one without the other (Seibt, 6 Jul 2025).

The cumulative picture is that minimal witnessing subsystems are best understood not as a single object class but as a general research pattern: identify the smallest or structurally sharpest subobject, support, zone, spectrum, or witness set that still certifies the property of interest. The technical content of that reduction depends on the semantic notion being preserved, the admissible form of reduction, and the chosen order of minimality.