Fidelity-Preserving Truncations

• Fidelity-preserving truncations are defined by retaining essential invariant structures—such as semantic closures, symmetry identities, and conservation rules—rather than preserving raw syntactic detail.
• They are applied across diverse fields including reversible causal computation, continuum QCD, tensor-network formulations, and prime-factor fibre analyses to guarantee task-specific fidelity.
• These schemes enable principled data reduction and compression while preserving critical system behavior, ensuring accurate rollback, symmetry preservation, and interval-specific performance.

Fidelity-preserving truncations are truncation schemes in which simplification is constrained by preservation of a distinguished invariant structure rather than by raw syntactic equality. In the recent literature, the preserved object may be a semantic closure relevant to rollback, exact Ward–Green–Takahashi identities and the matched relation between a gap equation and a Bethe–Salpeter kernel, local Kronecker-delta selection rules in tensor formulations, a small exceptional proportion on a conditioned prime-factor fibre, or the quadratic-output form together with accuracy on a prescribed time or frequency interval (Xu, 15 Jun 2026, Binosi et al., 2016, Meurice, 2019, Song et al., 2024). This suggests a cross-domain usage in which truncation is judged by preservation of semantics, symmetry, or task-relevant observables, not by literal retention of the original representation.

1. Conceptual profiles of fidelity

The literature does not use a single universal fidelity criterion. In reversible logging, fidelity is not syntactic equality of logs; it is semantic equality of closure, namely

$Cn_{\mathrm{rev}}(\hat S)=Cn_{\mathrm{rev}}(S_O).$

In the conditional Erdős–Wintner setting on prime-factor fibres, “fidelity-preserving” means that, on the relevant fibre, truncation changes $f$ on only a small proportion of integers. In tensor formulations of the nonlinear $O(2)$ sigma model and the compact Abelian Higgs model, tensor-network truncations can be made fidelity-preserving if they respect the exact symmetry structure already built into the tensor formulation. In balanced truncation for linear systems with quadratic outputs, the fidelity-preserving goal is not just small error, but small error in a prescribed interval while retaining the same quadratic-output form in the reduced-order model (Xu, 15 Jun 2026, Verwee, 18 Mar 2026, Meurice, 2019, Song et al., 2024).

These criteria are structurally different. Some are exact invariance conditions, such as closure preservation or exact symmetry identities. Others are controlled-loss conditions, such as an exceptional-set estimate or interval-focused approximation. A common misconception is that truncation fidelity is always equivalent to preserving the original data structure term by term. The reversible-logging framework makes the opposite point explicitly: many recorded facts are not essential; they are derivable from others under the reversible semantics, and deleting them does not change rollback behavior (Xu, 15 Jun 2026).

2. Closure-preserving truncation in reversible causal computation

In "Rate-Distortion for Reversible Causal Nets under Closure-Preserving Fidelity" (Xu, 15 Jun 2026), an execution history is modeled as a finite fact base

$S_O \subseteq \mathbb S,$

where $\mathbb S$ is the finite universe of all ground atoms over a reversible signature. Facts include dynamic history facts such as $\mathsf{Occurs}(e)$ or $\mathsf{In}(t)$, plus static rollback structure such as $\mathsf{Cause}$, $\mathsf{Conflict}$, and $\mathsf{Prevent}$. Semantics is represented by a monotone closure operator

$f$0

typically induced by a function-free Horn/Datalog system. Under Horn semantics, $f$1 satisfies extensivity, monotonicity, and idempotence. The paper’s “fidelity-preserving truncations” are the central compression idea: instead of preserving the raw log syntactically, the encoder keeps only a canonical irredundant core that preserves the semantic closure relevant to rollback.

The paper defines a closure-based similarity

$f$2

and a bounded single-symbol distortion

$f$3

Thus distortion is semantic, not syntactic. A change is low-distortion if the induced closure barely changes, even if the atom itself changes completely. Zero distortion satisfies

$f$4

The truncation mechanism is a deterministic deletion scan under a public total order $f$5 on $f$6. Starting with $f$7, one scans atoms $f$8 in order and deletes $f$9 if

$O(2)$0

The redundant remainder is

$O(2)$1

and the scan yields the decomposition

$O(2)$2

The theorem proved is that $O(2)$3 is uniquely determined by $O(2)$4 and every $O(2)$5 is irredundant:

$O(2)$6

This is a fidelity-preserving truncation because removing $O(2)$7 preserves closure exactly.

The rate-distortion consequences are unusually strong. Under the admissibility assumption

$O(2)$8

every redundant fact $O(2)$9 has zero distortion against every admissible reconstruction:

$S_O \subseteq \mathbb S,$0

Hence $S_O \subseteq \mathbb S,$1 is “invisible” under admissible reconstruction, and the full semantic rate-distortion problem collapses to the core:

$S_O \subseteq \mathbb S,$2

At perfect fidelity, overlaps among zero-distortion reconstruction sets

$S_O \subseteq \mathbb S,$3

induce the confusability hypergraph

$S_O \subseteq \mathbb S,$4

and the exact perfect-fidelity rate is

$S_O \subseteq \mathbb S,$5

The framework is instantiated on reversible causal nets and reversible prime event structures under multiple reversing disciplines. For causal and cause-respecting disciplines, the core is the frontier,

$S_O \subseteq \mathbb S,$6

where $S_O \subseteq \mathbb S,$7 is the set of maximal events under causal reachability. For inverse-causal discipline, the core may be larger than the frontier, because causes themselves must be retained to preserve blocker evidence. The paper’s practical conclusion is explicit: fidelity-preserving truncations are canonical semantic reductions of logs that preserve the closure relevant to rollback, enabling principled compression and rate-optimal reversible debugging.

3. Symmetry-preserving truncation in continuum QCD

In continuum QCD bound-state theory, fidelity-preserving truncation appears under the more established label of symmetry-preserving truncation. "Symmetry preserving truncations of the gap and Bethe-Salpeter equations" (Binosi et al., 2016) states that Ward–Green–Takahashi identities impose stringent relations between the kernel of the gap equation and the kernel of the Bethe–Salpeter equation. The canonical symmetry-preserving relation is

$S_O \subseteq \mathbb S,$8

The central message is that symmetry preservation in continuum QCD bound-state studies is not guaranteed by “improving” the quark–gluon vertex alone. What matters is the matched pair: the quark gap equation and the meson Bethe–Salpeter kernel must be constructed so that the relevant WGT identities are satisfied exactly.

The paper shows that naive dressed ladder constructions fail. A ladder-like kernel with one dressed vertex,

$S_O \subseteq \mathbb S,$9

or with both vertices dressed,

$\mathbb S$0

is generally insufficient. Using the Ball–Chiu Ansatz, the paper shows that the dressed ladder kernel $\mathbb S$1 can be tuned to produce $\mathbb S$2 in the chiral limit, but it still violates the axial-vector WGT identity. The diagnostic relation

$\mathbb S$3

fails because the ratio $\mathbb S$4 is not constant. The resulting correction to a common misconception is explicit: massless pion $\mathbb S$5 symmetry preservation.

The same paper identifies the origin and role of $\mathbb S$6-diagrams. Once the dressed gluon–quark vertex is represented in terms of the gluon–quark scattering matrix $\mathbb S$7, the differentiation $\mathbb S$8 generates not only ladder corrections and crossed-box-like terms, but also a class of two-loop, non-Abelian contributions involving the three-gluon vertex. These $\mathbb S$9-diagrams are two-particle-irreducible in the quark–antiquark channel, required by the symmetry-consistent differentiation of the gap equation, and cannot be absorbed into a dressed vertex or reproduced by adding crossed boxes. Therefore, there are no general circumstances under which the WGT identities can be preserved by a Bethe–Salpeter kernel obtained simply by dressing both gluon–quark vertices in a ladder-like truncation.

"Derivation of the gap and Bethe-Salpeter equations at large $\mathsf{Occurs}(e)$0 limit and symmetry preserving truncations" (Fu et al., 2017) develops the same principle from a common generating functional with a bilocal auxiliary field. There the kernel is obtained directly from the quark self-energy,

$\mathsf{Occurs}(e)$1

and the truncation scheme is organized order-by-order in the number of connected gluon legs appearing in the effective action. Because the quark gap equation, the quark–gluon vertex equation, and the meson Bethe–Salpeter equation are all derived from the same generating functional, truncating the functional automatically truncates all these equations in a mutually consistent way. The paper states explicitly that truncations of the generating functional preserve any linearly realized symmetry retained in the truncated functional; in particular, chiral symmetry in the chiral limit is preserved automatically. At leading order the scheme reduces to Rainbow-Ladder, whereas 3-gluon and 4-gluon connected Green’s functions generate non-Abelian corrections and H-shape Bethe–Salpeter diagrams. In this setting, fidelity preservation means maintaining the identity between the quark self-energy and the meson kernel required by chiral symmetry.

4. Tensor-network truncation and exact lattice constraints

"Examples of symmetry-preserving truncations in tensor field theory" (Meurice, 2019) studies the tensor formulation of the nonlinear $\mathsf{Occurs}(e)$2 sigma model and the compact Abelian Higgs model on a $\mathsf{Occurs}(e)$3-dimensional cubic lattice. The main truncation strategy is to keep only tensor indices with $\mathsf{Occurs}(e)$4, or similarly for plaquette/link quantum numbers, and set tensor elements with larger magnitude indices to zero. The key claim is that this is safe provided the truncation does not alter the exact Kronecker-delta constraints that encode charge/current conservation or gauge constraints. The truncation removes large-index dynamical weights, not the selection rules.

For the $\mathsf{Occurs}(e)$5 model, the action is

$\mathsf{Occurs}(e)$6

with global shift symmetry $\mathsf{Occurs}(e)$7. Global symmetry implies

$\mathsf{Occurs}(e)$8

hence if

$\mathsf{Occurs}(e)$9

then

$\mathsf{In}(t)$0

After Fourier expansion of each local Boltzmann factor, the tensor formulation contains the local tensor

$\mathsf{In}(t)$1

with

$\mathsf{In}(t)$2

The important part is the Kronecker delta, which is the discrete current conservation law. Because truncation leaves this delta constraint unchanged, charge conservation remains exact, forbidden correlators remain forbidden, and symmetry-based identities remain true.

The compact Abelian Higgs model is treated analogously. The matter and gauge actions are

$\mathsf{In}(t)$3

$\mathsf{In}(t)$4

with local gauge transformation

$\mathsf{In}(t)$5

After character expansion and integration over gauge fields, the link variables are constrained by a lattice Bianchi/continuity-type identity, written schematically as

$\mathsf{In}(t)$6

The paper’s conclusion is that gauge-variant one-point functions still vanish, gauge-invariant observables like Wilson loops remain allowed, and the local gauge constraint is not broken by discarding large-$\mathsf{In}(t)$7 tensor elements. A truncation would only become symmetry-breaking if one introduced tensors that violate the Kronecker-delta conservation law.

The significance is practical as well as formal. The paper argues that these truncations are desirable for implementations with quantum computers or for quantum simulations experiments because they produce a small local Hilbert space while retaining the exact symmetry constraints that define the physics.

5. Linear truncation on conditioned prime-factor fibres

"Linear truncation for conditioned prime-factor fibres" (Verwee, 18 Mar 2026) studies a real strongly additive function $\mathsf{In}(t)$8 and its truncation $\mathsf{In}(t)$9 in the conditional Erdős–Wintner theorem on the fibre $\mathsf{Cause}$0. The truncation is defined on prime powers by

$\mathsf{Cause}$1

and extended additively. Thus $\mathsf{Cause}$2 only modifies $\mathsf{Cause}$3 on the “bad” large primes

$\mathsf{Cause}$4

Since for $\mathsf{Cause}$5, $\mathsf{Cause}$6, the paper obtains

$\mathsf{Cause}$7

In this setting, “fidelity-preserving” means that, on the relevant fibre, truncation changes $\mathsf{Cause}$8 on only a small proportion of integers, so the distribution of $\mathsf{Cause}$9 on the fibre is well-approximated by the distribution of $\mathsf{Conflict}$0. The paper proves an effective linear truncation lemma in the central window

$\mathsf{Conflict}$1

under an effective Sathe–Selberg-type ratio estimate. The theorem states that

$\mathsf{Conflict}$2

uniformly for $\mathsf{Conflict}$3, $\mathsf{Conflict}$4 in the central window, and $\mathsf{Conflict}$5. For the classical fibre

$\mathsf{Conflict}$6

this becomes

$\mathsf{Conflict}$7

The improvement over earlier work is explicit. The old truncation term in the continuous case was of size

$\mathsf{Conflict}$8

whereas the new bound is on the natural linear scale

$\mathsf{Conflict}$9

The proof mechanism is direct. If $\mathsf{Prevent}$0, then $\mathsf{Prevent}$1 has at least one prime divisor $\mathsf{Prevent}$2. Writing

$\mathsf{Prevent}$3

and using stability of the fibre under removing one prime factor, one obtains

$\mathsf{Prevent}$4

The ratio estimate then converts one removed prime into a factor $\mathsf{Prevent}$5. The paper’s conclusion is that truncation becomes genuinely fidelity-preserving: it alters only a proportion $\mathsf{Prevent}$6 of the fibre, which is the natural linear scale expected from the harmonic density of the bad primes.

6. Structure-preserving and abstract relatives

"Balanced Truncation of Linear Systems with Quadratic Outputs in Limited Time and Frequency Intervals" (Song et al., 2024) develops a structure-preserving balanced truncation framework for stable linear time-invariant systems with quadratic outputs,

$\mathsf{Prevent}$7

and its multi-output extension. After a Petrov–Galerkin projection with $\mathsf{Prevent}$8 and $\mathsf{Prevent}$9, the reduced model satisfies

$f$00

so the reduced quadratic-output system preserves the same algebraic structure. Time-limited balanced truncation uses

$f$01

while frequency-limited balanced truncation uses

$f$02

with the corresponding limited-interval observability Gramian. The point is interval-specific fidelity: the most important states are those that are controllable and observable within the prescribed interval, not globally. The paper does not derive a new explicit TLBT error bound for the quadratic-output case; the guarantee is interval-focused accuracy, inherited by using interval-limited Gramians. A related practical point is that low-rank ADI and truncated Laguerre expansions are developed to make the method effective for large-scale systems.

A neighboring literature uses the term consistent truncation rather than fidelity-preserving truncation. "Consistent Truncations and Dualities" (Butter et al., 2022) defines a consistent truncation as a reduction from a higher-dimensional theory to a lower-dimensional one such that every solution of the reduced theory uplifts to a solution of the full theory. The paper proves that every dressing coset $f$03 yields a consistent truncation with generalized structure group $f$04, using the criterion that if $f$05 admits a generalized $f$06-structure with invariant tensors and constant singlet intrinsic torsion, then expanding all bosonic fields in terms of those invariant tensors yields a consistent truncation. A plausible implication is that this literature supplies a neighboring notion of truncation faithfulness, expressed as solution-level consistency rather than semantic closure, symmetry identity, or interval-restricted approximation.

At a more abstract level, "An Elementary Approach to Truncations" (Rasekh, 2018) studies truncation functors in an elementary $f$07-topos. It distinguishes external truncation, defined by the condition that for every object $f$08, $f$09 is $f$10-truncated, from internal truncation, defined using universes, the natural number object, and internal spheres. The internal criterion is

$f$11

The paper then constructs localization functors from ideal subuniverses via an internal right Kan extension and uses this machinery to construct truncation functors $f$12. This suggests an abstract reformulation of truncation in terms of universal properties and localization, rather than only in terms of data deletion or approximation.

Across these literatures, fidelity-preserving truncation is therefore not a single algorithmic template. It is a family of truncation doctrines in which the admissible reduction is defined by what must remain exact: closure relevant to rollback, matched symmetry identities, exact tensor selection rules, conditioned-fibre distributional behavior, quadratic-output structure on a specified interval, or a universal localization property.