Pure Variable Assumption in Theory & Practice

• Pure Variable Assumption is a conceptual framework that isolates a key variable or component to prevent confounding, thereby enhancing clarity and interpretability across theoretical models.
• It is applied in diverse fields—from enforcing invariance in Bell inequality tests and achieving assumption-lean inference in statistics to identifying latent structures in causal graphs and quantum state constructions.
• The approach often substitutes fragile model-based quantities with structured surrogates (e.g., double-triangular conditions or overlap-weighted contrasts) to maintain robust interpretability without collapsing essential model context.

Searching arXiv for the cited works and closely related material on the term and its usages. In the cited literature, the “Pure Variable Assumption” does not denote a single formally standardized object. As an Editor’s term, it names a recurrent methodological move: a theory, estimand, graphical condition, or state description is organized so that one variable, one interaction term, one latent parent, or one pure-state component is isolated from confounding structure that would otherwise obscure interpretation. In some works this isolation is treated critically, as in a claimed hidden assumption behind Bell inequalities; in others it is the explicit design goal of assumption-lean inference, identifiability theory, constrained argumentation, or quantum-state constructions. The resulting family of meanings is therefore heterogeneous rather than canonical (Lokajíček, 2011, Vansteelandt et al., 2020, Lee et al., 23 May 2025, Farah, 10 Mar 2025, Anwar et al., 2014, Oda, 2011).

1. Bell-type interchange invariance and the classicalization critique

In Lokajíček’s analysis of Bell inequalities, the relevant assumption is not presented as a generic appeal to “local realism,” but as a specific invariance requirement on Bell’s probability combination

$B=a_1b_1+a_2b_1+a_1b_2-a_2b_2\le 2,$

where $a_j$ and $b_k$ are transmission probabilities for two photons passing through polarizers at different angles. The paper states that Bell assumed the value of $B$ for a given combination of four individual probabilities remained unchanged when one pair of probabilities in one polarizer was interchanged. The author interprets this as requiring coincidence probabilities practically not to depend on spin orientations, and therefore as a restriction that is already classical in character (Lokajíček, 2011).

That critique is then recast in operator language. With $a_j$ and $b_k$ treated as operators on $H=H_a\oplus H_b$, the paper distinguishes three regimes by commutation structure. A noncommuting regime,

$$[a_j,b_k]\neq 0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$

is associated with possible interaction at distance. A hidden-variable local regime,

$$[a_j,b_k]=0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$

retains quantum characteristics at the level of individual polarizers while excluding direct long-distance interaction. A fully classical regime,

$[a_j,b_k]=0,\quad [a_1,a_2]=[b_1,b_2]=0,$

is identified as deterministic and free of quantum characteristics. The paper’s central claim is that Bell’s derivation effectively selects the third regime, so that experimental violations near $a_j$0 exclude only that fully commuting classical alternative, not all hidden-variable theories. This position is explicitly aligned with Einstein’s ontological rejection of nonlocality and with the view that there is no reason to believe in direct interaction between matter objects at great distances (Lokajíček, 2011).

A common misconception addressed by this line of argument is that Bell-inequality violation, by itself, refutes every locality-preserving hidden-variable model. The paper disputes that conclusion by locating the decisive force in the interchange-invariance requirement rather than in a universally neutral test of all hidden-variable structures. This is the paper’s own interpretation, not a field-wide terminological standard.

2. Assumption-lean statistics, instrumental variables, and growth assumptions

In semiparametric statistics, the corresponding “purity” idea is constructive rather than critical. For generalized linear models, Vansteelandt and Dukes redefine main effects and effect modification effects as nonparametric estimands that remain scientifically interpretable even when the working GLM is misspecified. For the generalized partially linear model

$a_j$1

the main-effect estimand is defined by

$a_j$2

and for binary $a_j$3 it becomes an overlap-weighted average of within-$a_j$4 contrasts. Under correct specification it reduces to the usual GLM coefficient; under misspecification it remains a weighted summary of the conditional association between $a_j$5 and $a_j$6. The same logic is applied to interaction terms, including the binary-$a_j$7 overlap-weighted contrast and, for continuous exposures, an orthogonal-projection definition using $a_j$8. Estimation proceeds through efficient influence curves under the full nonparametric model, with sample-splitting, cross-fitting, sandwich variance, and Wald intervals; the stated implementation for the main effect is to estimate $a_j$9, $b_k$0, and $b_k$1, form $b_k$2, regress $b_k$3 on $b_k$4 without intercept, and preferably use 10-fold cross-fitting (Vansteelandt et al., 2020).

In instrumental-variable analysis, the analogous assumption is the effective random assignment, or as-if randomization, of the instrument $b_k$5 conditional on observed covariates $b_k$6. The paper formalizes the null as

$b_k$7

and proposes an exact nonparametric randomization test: specify $b_k$8, choose a balance or bias statistic $b_k$9, simulate assignments, and compute a randomization-based $B$0-value. The method permits both covariate-specific diagnostics and global balance measures such as the Mahalanobis distance

$B$1

In the SPOTlight ICU application with 13,011 deteriorating ward patients in 48 NHS hospitals, ICU bed availability was better balanced than ICU admission itself; under complete randomization neither instrument nor exposure matched the benchmark, but under Bernoulli-trial assignment models the instrument was substantially closer to complete randomization than the exposure (Branson et al., 2019).

A third statistical-algorithmic usage appears in stochastic optimization, where the supplied commentary associates the label with the classical Blum–Gladyshev growth condition

$B$2

The paper treats this as a weak variance-type assumption, weaker than bounded stochastic subgradient norms and more general than bounded-variance assumptions tailored to smooth SGD. Its main technical device is Halpern anchoring,

$B$3

which introduces a negative quadratic term capable of canceling the growth term coming from the oracle bound. The stated consequences include horizon-free, anytime guarantees, a weighted-average $B$4 rate, a last-iterate $B$5 guarantee, and extensions to functional constraints and convex-concave min-max problems without bounded feasible sets (Alacaoglu et al., 14 Apr 2025).

These three usages share a common structure only at a high level. Each replaces a fragile model-based quantity with an object intended to survive misspecification, confounding, or unbounded growth. They do not define a single unified technical assumption.

3. Latent-variable identifiability beyond pure children

In latent causal graphical models, the relevant purity notion is the classical pure-child assumption: each latent variable has one or more observed children with exactly one latent parent. The 2025 identifiability paper relaxes that requirement in a Binary Latent Causal Model $B$6, $B$7, with arbitrary observed-variable types, binary latent variables, no observed-to-latent edges, no edges among observed variables, full-graph causal Markov and faithfulness conditions, and nondegeneracy conditions including positive probability for all latent configurations and at least one observed child for each latent (Lee et al., 23 May 2025).

The replacement is the double triangular condition on the bipartite adjacency matrix $B$8. A $B$9 binary matrix is triangular if, after row and column permutations, it has ones on the diagonal and arbitrary $a_j$0 entries below the diagonal. The matrix $a_j$1 is double triangular if, after row permutation,

$a_j$2

with $a_j$3 and $a_j$4 triangular $a_j$5 blocks and $a_j$6 arbitrary. Setting all starred entries in $a_j$7 to zero yields $a_j$8, which is exactly the condition of two pure children per latent variable. The double-triangular condition therefore strictly generalizes the pure-child setting (Lee et al., 23 May 2025).

The identifiability results are correspondingly stronger than the classical pure-child template. If the true $a_j$9 is double triangular, then the number of latent variables $b_k$0 is identifiable. Assuming known $b_k$1, if $b_k$2 is double triangular and all columns in $b_k$3 are nonempty, then $b_k$4 itself is identifiable. If, in addition, the subset condition holds,

$b_k$5

for any distinct $b_k$6, then the latent graph $b_k$7 and latent distribution $b_k$8 are identifiable; a monotonicity condition provides an alternative route that also resolves sign-flip ambiguity. The proofs use full-column-rank conditional probability tables induced by triangular blocks and a Kruskal-unique tensor decomposition of $b_k$9. The paper is also explicit about limitations: binary latents, exclusion of hidden variables without observed children, and a gap between sufficient and necessary conditions (Lee et al., 23 May 2025).

A central misconception corrected here is that pure children are necessary for nonparametric latent identifiability. The paper’s conclusion is precisely the opposite: what is necessary is not purity in the sense of isolated children, but enough structured asymmetry in the measurement graph.

4. Local one-coordinate dependence and schematic assumptions

In combinatorics, “pure variable” behavior is formalized as local dependence on a single coordinate. For a function

$H=H_a\oplus H_b$0

the theorem in “Dependence of functions on their variables” states that exactly one of two alternatives holds. Either each $H=H_a\oplus H_b$1 admits a finite partition $H=H_a\oplus H_b$2 such that on every cell $H=H_a\oplus H_b$3 the function factors through one coordinate projection $H=H_a\oplus H_b$4, or there exists a partition $H=H_a\oplus H_b$5 into nonempty sets and sequences

$H=H_a\oplus H_b$6

such that for all $H=H_a\oplus H_b$7 and all $H=H_a\oplus H_b$8,

$H=H_a\oplus H_b$9

The second alternative is the exact obstruction to local one-variable dependence. The paper proves that the two alternatives are mutually exclusive, derives the dichotomy by ultrafilter and compactness arguments, shows that the full theorem is not provable in ZF by a Dedekind-finite counterexample, and proves a positive ZF result for well-orderable sets (Farah, 10 Mar 2025).

In constrained assumption-based argumentation, the departure from “purely ground” reasoning takes a different form. CABA frameworks

$$[a_j,b_k]\neq 0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$0

allow variables and constraints in assumptions, rules, and arguments. Contraries are uniform across instantiations: if $$[a_j,b_k]\neq 0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$1 and $$[a_j,b_k]\neq 0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$2 are assumptions, one contrary predicate symbol $$[a_j,b_k]\neq 0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$3 satisfies

$$[a_j,b_k]\neq 0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$4

Rules may be non-ground and constraint-bearing, tight constrained arguments are tree-structured derivations whose leaves may include both assumptions and constraints, and every tight constrained argument is an instance of a most general constrained argument. Non-ground semantics are defined through full, partial, and precise attack, and the framework is proved to conservatively generalize standard ABA through the grounding construction $$[a_j,b_k]\neq 0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$5 (Angelis et al., 13 Feb 2026).

These two works are closely aligned at the level of structure. The combinatorial theorem says that a function is either piecewise one-coordinate or exhibits a strong obstruction; CABA says that a variable-bearing assumption schema can stand for infinitely many ground instances without reducing semantics to a purely variable-free level. In both cases, “purity” is local and schematic, not absolute.

5. Quantum information: pure states, restricted measurements, and mixed-state computation

In multipartite quantum information, one purity assumption concerns the measurement set rather than the state space as a whole. The PEPS paper fixes a restricted local measurement set $$[a_j,b_k]\neq 0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$6, defines its dual

$$[a_j,b_k]\neq 0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$7

and calls operators strictly $$[a_j,b_k]\neq 0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$8-positive when all inequalities are strict. Its key assumption is that $$[a_j,b_k]\neq 0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$9 strictly contains a pure quantum state, that is, contains an open neighborhood of some rank-1 projector $$[a_j,b_k]=0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$0. Under this condition, and with $$[a_j,b_k]=0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$1, the authors construct pure entangled PEPS that are $$[a_j,b_k]=0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$2-separable, hence admit local hidden-variable models for measurements from $$[a_j,b_k]=0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$3; in a second recipe, the hidden-variable distribution factorizes so that sampling is efficient. The same states are not classical without the measurement restriction, and in some proof-of-principle cases are universal for measurement-based quantum computation (Anwar et al., 2014).

A nearly opposite use of the purity concept appears in continuous-variable computation without cooling. There the paper rejects the conventional assumption that logical states must be encoded in pure physical states prepared near the ground state. Instead, logical information is encoded in mixed physical states via parity structure. The proposed two-qumode parity encoding uses basis pairs

$$[a_j,b_k]=0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$4

supports universal logical gates $$[a_j,b_k]=0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$5, $$[a_j,b_k]=0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$6, and $$[a_j,b_k]=0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$7, and is implemented with auxiliary-qubit-mediated parity and swap-like operations built from realistic hybrid interactions. The stated consequence is that ground-state cooling is no longer a prerequisite: thermal mixed states can be parity-projected into logical states, with lower initialization energy and resilience to collective phase-shift and collective squeezing noise (Lau et al., 2016).

State certification and verification supply a third quantum usage. The heterodyne-based protocol for continuous-variable states introduces a direct estimator $$[a_j,b_k]=0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$8 for expectation values of bounded-support operators from Husimi-$$[a_j,b_k]=0,\quad [a_1,a_2]\neq 0,\quad [b_1,b_2]\neq 0,$$9 samples. For tomography on i.i.d. bounded-support states, this yields matrix-element estimates with analytical confidence intervals and no reconstruction or binning. For verification, the target state is pure, but the tested state is not assumed pure, i.i.d., or honestly prepared: the prover may be fully malicious and the global state arbitrarily entangled across subsystems. The protocol combines support estimation, a de Finetti reduction, and Hoeffding-type concentration to bound the fidelity estimator for $[a_j,b_k]=0,\quad [a_1,a_2]=[b_1,b_2]=0,$0 retained systems. A crucial distinction is therefore enforced between purity of the target and purity of the source (Chabaud et al., 2019).

Across these quantum settings, purity is redistributed rather than universally imposed. It may reside in an interior point of a dual measurement cone, in the target of a verification problem, or in the logical action of gates on a parity-defined subsystem, while the physical state itself may be entangled, measurement-restricted, or mixed.

6. Pure spinor as the sole fundamental variable

In Berkovits-type string theory, the pure-variable idea is literal. The starting point is a classical action whose only fundamental dynamical variable is the bosonic pure spinor $[a_j,b_k]=0,\quad [a_1,a_2]=[b_1,b_2]=0,$1, constrained by

$[a_j,b_k]=0,\quad [a_1,a_2]=[b_1,b_2]=0,$2

For the superparticle,

$[a_j,b_k]=0,\quad [a_1,a_2]=[b_1,b_2]=0,$3

and for the superstring,

$[a_j,b_k]=0,\quad [a_1,a_2]=[b_1,b_2]=0,$4

The classical theory is topological: its constraints are first class, and the action is trivial on-shell (Oda, 2011).

BRST gauge fixing then generates the standard superspace variables as ghosts. In the superparticle, the fermionic Faddeev–Popov ghost is reinterpreted as $[a_j,b_k]=0,\quad [a_1,a_2]=[b_1,b_2]=0,$5, while reducibility of the symmetry introduces bosonic ghosts-for-ghosts $[a_j,b_k]=0,\quad [a_1,a_2]=[b_1,b_2]=0,$6, parameterized as

$[a_j,b_k]=0,\quad [a_1,a_2]=[b_1,b_2]=0,$7

This leads to the BRST charge

$[a_j,b_k]=0,\quad [a_1,a_2]=[b_1,b_2]=0,$8

with

$[a_j,b_k]=0,\quad [a_1,a_2]=[b_1,b_2]=0,$9

and to the action

$a_j$00

For the superstring, the same logic yields

$a_j$01

with

$a_j$02

together with the standard BRST variations

$a_j$03

The paper’s interpretation is that the pure spinor formalism may originate in a topological field theory, with $a_j$04 and $a_j$05 emerging only after BRST quantization (Oda, 2011).

This is the most ontologically stringent version of the Pure Variable Assumption in the supplied corpus. The variable is not merely isolated for analysis; it is treated as the sole fundamental degree of freedom from which the rest of the formalism emerges.

7. Conceptual status and recurrent misunderstandings

Several recurrent misunderstandings are corrected by the cited works. First, the phrase should not be treated as a single cross-disciplinary axiom. Some papers do not define it formally at all; some use only the adjacent notions of pure state, pure children, or pure spinor; and one paper explicitly notes that it does not introduce a separate technical notion of “pure variable” distinct from pure state (Lokajíček, 2011, Anwar et al., 2014).

Second, “purity” rarely means absence of all contextual structure. In the Bell paper, the disputed issue is invariance under interchange of detector probabilities; in assumption-lean regression, the goal is a nonparametric estimand that continues to summarize conditional association under misspecification; in latent graphical models, pure children are only one sufficient route to identifiability and are not necessary; in CABA, variable-bearing assumptions are preserved rather than eliminated; in continuous-variable quantum computing, mixed physical states can still support universal logic (Vansteelandt et al., 2020, Lee et al., 23 May 2025, Angelis et al., 13 Feb 2026, Lau et al., 2016).

Third, a purity assumption is often replaced by a structural surrogate. Double-triangular measurement graphs replace pure children; overlap-weighted contrasts replace model-correctness assumptions; randomization tests replace unexamined acceptance of as-if random assignment; heterodyne estimators replace tomography by unstable reconstruction; Halpern anchoring replaces bounded-variance-type regularity with a growth condition that is directly controlled in the iteration (Lee et al., 23 May 2025, Branson et al., 2019, Chabaud et al., 2019, Alacaoglu et al., 14 Apr 2025).

A plausible synthesis is that the “Pure Variable Assumption” marks a boundary question rather than a single doctrine: when may one attribute a phenomenon to one variable, one effect, one latent source, or one pure component without collapsing the surrounding structure that gives the phenomenon its real content? The cited literature gives sharply different answers, but it is unified in treating that boundary as mathematically substantive rather than terminological.