BLCE in Multidisciplinary Research

• BLCE is an overloaded acronym representing diverse constructs such as constrained linear estimation, blur-adaptive camera trajectory recovery, bandit elimination methods, neutrino mixing schemes, and post-common-envelope stellar evolution channels.
• In statistical estimation and online learning, BLCE leverages constrained optimization and rare-update elimination techniques to reduce variance and achieve near-minimax optimal performance.
• In astrophysics and neutrino phenomenology, BLCE models delicate phenomena—from CP-dependent lepton mixing to pulsation driving in BLAPs—highlighting the need for precise, discipline-specific calibration.

BLCE is an overloaded acronym used in several technically unrelated research areas. In the arXiv literature represented here, it denotes at least five distinct constructs: the Best Linear Constrained Estimator in statistical estimation theory, Blur-adaptive Latent Camera Estimation in dynamic novel view synthesis, Batched-Feedback Linear Contextual Bandit with Elimination in online learning, Bi-Large mixing with charged lepton correction in neutrino phenomenology, and a post-common-envelope evolutionary origin for blue large-amplitude pulsators in stellar astrophysics (Lang et al., 2017, Bui et al., 21 Apr 2025, Yu et al., 31 May 2026, Roy et al., 2012, Byrne et al., 2018). The acronym is therefore not a field-independent term, and its meaning is determined almost entirely by disciplinary context.

1. Estimation theory: constrained BLUE as BLCE

In estimation theory, BLCE denotes the Best Linear Constrained Estimator, i.e. the constrained version of the best linear unbiased estimator under linear equality constraints (Lang et al., 2017). The underlying real-valued model is

$y = Hx + w,$

with $y \in \mathbb{R}^{N_y}$, $x \in \mathbb{R}^{N_x}$, $H \in \mathbb{R}^{N_y\times N_x}$, $E[w]=0$, and $\mathrm{Cov}(w)=R \succ 0$. The parameter vector satisfies consistent linear constraints

$Cx = d,$

where $C \in \mathbb{R}^{N_c \times N_x}$ has full row rank.

When $H$ has full column rank, the unconstrained BLUE coincides with the GLS estimator,

$$\hat{x}_{\mathrm{BLUE}} = \bigl(H^\top R^{-1} H\bigr)^{-1} H^\top R^{-1} y,$$

with information matrix $y \in \mathbb{R}^{N_y}$0. The constrained estimator is obtained from the constrained GLS problem

$y \in \mathbb{R}^{N_y}$1

whose KKT system is

$y \in \mathbb{R}^{N_y}$2

The resulting closed form is

$y \in \mathbb{R}^{N_y}$3

or equivalently,

$y \in \mathbb{R}^{N_y}$4

A null-space formulation is also central. If $y \in \mathbb{R}^{N_y}$5, where $y \in \mathbb{R}^{N_y}$6 and $y \in \mathbb{R}^{N_y}$7, then under the weaker condition that $y \in \mathbb{R}^{N_y}$8 has full column rank,

$y \in \mathbb{R}^{N_y}$9

This form remains valid even when $x \in \mathbb{R}^{N_x}$0 is not full column rank or $x \in \mathbb{R}^{N_x}$1, provided $x \in \mathbb{R}^{N_x}$2 (Lang et al., 2017).

The constrained estimator is unbiased on the feasible affine subspace, $x \in \mathbb{R}^{N_x}$3 when $x \in \mathbb{R}^{N_x}$4, and its covariance is

$x \in \mathbb{R}^{N_x}$5

equivalently

$x \in \mathbb{R}^{N_x}$6

The constrained covariance satisfies

$x \in \mathbb{R}^{N_x}$7

so the constraints remove variability along constrained directions. Under white noise, $x \in \mathbb{R}^{N_x}$8, the BLCE reduces to the constrained least-squares estimator, which is one reason the constrained BLUE and constrained LS are algebraically so closely related (Lang et al., 2017).

2. Dynamic 3D Gaussian splatting: Blur-adaptive Latent Camera Estimation

In computer vision, BLCE denotes Blur-adaptive Latent Camera Estimation, the camera-trajectory module of MoBGS for deblurring dynamic 3D Gaussian Splatting from blurry monocular video (Bui et al., 21 Apr 2025). Its purpose is to recover a per-frame continuous latent camera trajectory during the exposure interval and thereby model blur induced by global camera motion.

The blurry frame $x \in \mathbb{R}^{N_x}$9 is modeled under global shutter and uniform exposure as

$H \in \mathbb{R}^{N_y\times N_x}$0

with discrete approximation

$H \in \mathbb{R}^{N_y\times N_x}$1

BLCE adapts trajectory complexity to blur severity. A blur score is computed from the low-frequency energy ratio of the shifted DFT magnitude,

$H \in \mathbb{R}^{N_y\times N_x}$2

This score is encoded as a blur feature

$H \in \mathbb{R}^{N_y\times N_x}$3

The latent trajectory is then parameterized in five steps: a seed latent code is formed from blur and pose features, that code is evolved by a neural ODE over $H \in \mathbb{R}^{N_y\times N_x}$4 exposure samples, each evolved state is decoded into a twist $H \in \mathbb{R}^{N_y\times N_x}$5, the twist is mapped to $H \in \mathbb{R}^{N_y\times N_x}$6 via the exponential map, and the residual transform is composed with the training pose: $H \in \mathbb{R}^{N_y\times N_x}$7

$H \in \mathbb{R}^{N_y\times N_x}$8

$H \in \mathbb{R}^{N_y\times N_x}$9

$E[w]=0$0

$E[w]=0$1

This construction is paired with Latent Camera-induced Exposure Estimation (LCEE), which couples exposure time to actual camera path length across static Gaussians: $E[w]=0$2 The estimated $E[w]=0$3 then determines the latent timestamps used for rendering dynamic object motion. This enforces the paper’s central physical consistency requirement: global camera blur and local object blur are integrated over the same exposure interval (Bui et al., 21 Apr 2025).

MoBGS optimizes BLCE jointly with static and dynamic Gaussian parameters, spline control points, and several losses: $E[w]=0$4 The reported implementation uses $E[w]=0$5, with experiments showing improvements up to $E[w]=0$6, reported training time of approximately $E[w]=0$7 h per scene, and rendering speed of approximately $E[w]=0$8 FPS. On Stereo Blur, MoBGS with BLCE/LCEE attains $E[w]=0$9 in PSNR/SSIM/LPIPS/tOF for the full region (Bui et al., 21 Apr 2025).

3. Online learning: Batched-Feedback Linear Contextual Bandit with Elimination

In contextual bandits, BLCE denotes Batched-Feedback Linear Contextual Bandit with Elimination, a rare-update algorithm for linear contextual bandits (Yu et al., 31 May 2026). The setting is

$\mathrm{Cov}(w)=R \succ 0$0

where $\mathrm{Cov}(w)=R \succ 0$1 is unknown, $\mathrm{Cov}(w)=R \succ 0$2 is independent $\mathrm{Cov}(w)=R \succ 0$3-sub-Gaussian noise, and each round reveals a set of $\mathrm{Cov}(w)=R \succ 0$4 context vectors

$\mathrm{Cov}(w)=R \succ 0$5

The regret is

$\mathrm{Cov}(w)=R \succ 0$6

The distinctive operational constraint is that reward-dependent parameter retraining is permitted only $\mathrm{Cov}(w)=R \succ 0$7 times. BLCE uses a static schedule with update times $\mathrm{Cov}(w)=R \succ 0$8, where $\mathrm{Cov}(w)=R \succ 0$9, the first interval has length approximately

$Cx = d,$0

and later interval lengths scale like

$Cx = d,$1

The total number of intervals is at most $Cx = d,$2 (Yu et al., 31 May 2026).

BLCE is explicitly not strictly-batched. Parameter estimates $Cx = d,$3 are updated only at interval boundaries, but within intervals the algorithm continues to maintain reward-free quantities such as Gram matrices and uses them in action selection. In the first interval, it performs pure uncertainty-driven exploration: $Cx = d,$4 with inverse Gram update by Sherman–Morrison,

$Cx = d,$5

At the first update,

$Cx = d,$6

with $Cx = d,$7.

For later intervals, BLCE begins with multi-level elimination. Given past estimates $Cx = d,$8, it defines nested feasible sets

$Cx = d,$9

$C \in \mathbb{R}^{N_c \times N_x}$0

It then splits the interval into an uncertainty-driven exploration phase and a greedy exploitation phase with respect to $C \in \mathbb{R}^{N_c \times N_x}$1. The confidence widths are

$C \in \mathbb{R}^{N_c \times N_x}$2

$C \in \mathbb{R}^{N_c \times N_x}$3

$C \in \mathbb{R}^{N_c \times N_x}$4

The main theoretical guarantee is

$C \in \mathbb{R}^{N_c \times N_x}$5

described as minimax-optimal up to polylogarithmic factors (Yu et al., 31 May 2026). Its total runtime is

$C \in \mathbb{R}^{N_c \times N_x}$6

and the paper emphasizes that BLCE attains this guarantee without the near G-optimal design step used by BLCE-G, thereby avoiding the $C \in \mathbb{R}^{N_c \times N_x}$7 per-call bottleneck associated with such designs (Yu et al., 31 May 2026).

4. Neutrino phenomenology: Bi-Large mixing with charged lepton correction

In neutrino physics, BLCE denotes Bi-Large mixing with charged lepton correction, a lepton-mixing construction in which the neutrino sector is assumed to obey Bi-Large mixing and the charged-lepton sector contributes a CKM-type correction (Roy et al., 2012).

The Bi-Large ansatz takes the Cabibbo angle as the seed parameter: $C \in \mathbb{R}^{N_c \times N_x}$8 with the strict BL choice

$C \in \mathbb{R}^{N_c \times N_x}$9

and $H$0. The neutrino mixing matrix is $H$1 in standard PDG form. The charged-lepton correction is a CKM-like $H$2-rotation,

$H$3

and the physical PMNS matrix is

$H$4

Because $H$5 in the Bi-Large ansatz, all three observable mixing angles depend on $H$6. The leading relations may be written as

$H$7

with analogous $H$8-dependent formulas for $H$9 and $$\hat{x}_{\mathrm{BLUE}} = \bigl(H^\top R^{-1} H\bigr)^{-1} H^\top R^{-1} y,$$0. The paper gives explicit closed forms: $$\hat{x}_{\mathrm{BLUE}} = \bigl(H^\top R^{-1} H\bigr)^{-1} H^\top R^{-1} y,$$1

$$\hat{x}_{\mathrm{BLUE}} = \bigl(H^\top R^{-1} H\bigr)^{-1} H^\top R^{-1} y,$$2

$$\hat{x}_{\mathrm{BLUE}} = \bigl(H^\top R^{-1} H\bigr)^{-1} H^\top R^{-1} y,$$3

A key motivation is to avoid the well-known limitation of Bi-Maximal mixing with charged-lepton correction. In the BM case, the minimum solar angle is $$\hat{x}_{\mathrm{BLUE}} = \bigl(H^\top R^{-1} H\bigr)^{-1} H^\top R^{-1} y,$$4, equivalently $$\hat{x}_{\mathrm{BLUE}} = \bigl(H^\top R^{-1} H\bigr)^{-1} H^\top R^{-1} y,$$5, and this minimum occurs only for $$\hat{x}_{\mathrm{BLUE}} = \bigl(H^\top R^{-1} H\bigr)^{-1} H^\top R^{-1} y,$$6, implying vanishing CP violation. In the BLCE construction, by contrast, a single choice of phase can align all three angles. For $$\hat{x}_{\mathrm{BLUE}} = \bigl(H^\top R^{-1} H\bigr)^{-1} H^\top R^{-1} y,$$7, the paper finds

$$\hat{x}_{\mathrm{BLUE}} = \bigl(H^\top R^{-1} H\bigr)^{-1} H^\top R^{-1} y,$$8

leading to

$$\hat{x}_{\mathrm{BLUE}} = \bigl(H^\top R^{-1} H\bigr)^{-1} H^\top R^{-1} y,$$9

The scheme also predicts that $y \in \mathbb{R}^{N_y}$00 lies in the first octant, with

$y \in \mathbb{R}^{N_y}$01

under variation of $y \in \mathbb{R}^{N_y}$02 (Roy et al., 2012).

The associated Jarlskog invariant is

$y \in \mathbb{R}^{N_y}$03

and for BLCE the paper derives

$y \in \mathbb{R}^{N_y}$04

This makes BLCE a highly constrained and therefore highly testable mixing scheme (Roy et al., 2012).

5. Stellar astrophysics: the BLCE evolutionary channel for BLAPs

In stellar astrophysics, BLCE refers to a post-common-envelope evolutionary origin for blue large-amplitude pulsators (BLAPs) (Byrne et al., 2018). The term identifies a channel in which a low-mass star loses its hydrogen envelope in a rapid common-envelope ejection and subsequently contracts through the BLAP region before becoming either a low-mass helium white dwarf or a core-helium-burning extreme horizontal-branch star.

The observational BLAP domain summarized in the models is

$y \in \mathbb{R}^{N_y}$05

with large-amplitude single-period photometric variations of about $y \in \mathbb{R}^{N_y}$06 mag on timescales of $y \in \mathbb{R}^{N_y}$07 minutes (Byrne et al., 2018). The post-CE calculations begin from a $y \in \mathbb{R}^{N_y}$08 red-giant progenitor that undergoes rapid envelope loss at

$y \in \mathbb{R}^{N_y}$09

leaving a thin hydrogen layer of approximately $y \in \mathbb{R}^{N_y}$10. Two remnant masses were modeled: approximately $y \in \mathbb{R}^{N_y}$11, which ignites helium and approaches the EHB, and approximately $y \in \mathbb{R}^{N_y}$12, which contracts without helium ignition and becomes a low-mass helium white dwarf (Byrne et al., 2018).

The decisive microphysics is atomic diffusion, particularly radiative levitation of iron-group elements. Radiative accelerations satisfy

$y \in \mathbb{R}^{N_y}$13

and diffusion velocities schematically obey

$y \in \mathbb{R}^{N_y}$14

These processes lead to Fe/Ni accumulation near the iron-group opacity bump at

$y \in \mathbb{R}^{N_y}$15

With levitation, the enhanced Fe/Ni opacity bump drives the fundamental radial mode through the $y \in \mathbb{R}^{N_y}$16-mechanism. Stability is measured by the work integral

$y \in \mathbb{R}^{N_y}$17

with instability corresponding to $y \in \mathbb{R}^{N_y}$18 or $y \in \mathbb{R}^{N_y}$19. The data explicitly state that without diffusion the Z-bump is too weak and the fundamental mode is stable, whereas with levitation the local Fe/Ni enhancement increases the driving term by roughly an order of magnitude and renders the surface value of $y \in \mathbb{R}^{N_y}$20 positive in the BLAP domain (Byrne et al., 2018).

The predicted fundamental-mode periods are consistent with observation: $y \in \mathbb{R}^{N_y}$21 and for $y \in \mathbb{R}^{N_y}$22 and $y \in \mathbb{R}^{N_y}$23,

$y \in \mathbb{R}^{N_y}$24

Both sequences pass through the BLAP $y \in \mathbb{R}^{N_y}$25 box, but the $y \in \mathbb{R}^{N_y}$26 sequence is described as the most consistent with BLAPs. Its period-change magnitude is

$y \in \mathbb{R}^{N_y}$27

with negative sign due to contraction, matching the order of many observed BLAPs. The $y \in \mathbb{R}^{N_y}$28 pre-EHB sequence produces $y \in \mathbb{R}^{N_y}$29 about an order of magnitude too large, and neither contracting sequence naturally explains positive $y \in \mathbb{R}^{N_y}$30, except through speculative brief expansion loops during off-center He ignition (Byrne et al., 2018).

A related paper presents the same evolutionary picture with further numerical detail, including representative unstable models at $y \in \mathbb{R}^{N_y}$31, $y \in \mathbb{R}^{N_y}$32, $y \in \mathbb{R}^{N_y}$33 for the $y \in \mathbb{R}^{N_y}$34 case and $y \in \mathbb{R}^{N_y}$35, $y \in \mathbb{R}^{N_y}$36, $y \in \mathbb{R}^{N_y}$37 for the $y \in \mathbb{R}^{N_y}$38 case (Byrne et al., 2018). This suggests that, within the BLCE astrophysical usage, the acronym refers not to a named algorithm but to a specific post-common-envelope channel for BLAP phenomenology.

6. Ambiguity, nonstandard usage, and related acronym collisions

BLCE is not uniformly standardized across the cited literature. Some papers explicitly note that the acronym is absent or unofficial in their own setting. In ML-assisted communications, for example, the paper on BER- and BLER-oriented training states that the paper does not define the acronym BLCE; instead it introduces block-level surrogates such as Product, SmoothMax, LogSumExp, and $y \in \mathbb{R}^{N_y}$39-norm losses for minimizing BLER directly (Wiesmayr et al., 2022). Any reading of BLCE in that context as a named method would therefore be inaccurate.

A similar caution applies in continual learning. The paper “Bayesian Learning-driven Prototypical Contrastive Loss for Class-Incremental Learning” names its method BLCL, not BLCE (Raichur et al., 2024). The details note that if one encounters BLCE as “Bayesian learning-driven contrastive approach balancing cross-entropy,” it is only a conceptual mapping to the same uncertainty-based weighting between cross-entropy and contrastive losses: $y \in \mathbb{R}^{N_y}$40 The official notation in that paper remains BLCL (Raichur et al., 2024).

This multiplicity of meanings has a practical implication for literature search and citation. In statistics and signal processing, BLCE is tied to constrained Gauss–Markov estimation (Lang et al., 2017); in dynamic 3DGS it is a latent camera estimation module (Bui et al., 21 Apr 2025); in bandit theory it is a rare-update elimination algorithm (Yu et al., 31 May 2026); in neutrino phenomenology it is a mixing ansatz with charged-lepton correction (Roy et al., 2012); and in BLAP studies it abbreviates a post-common-envelope evolutionary channel (Byrne et al., 2018). The acronym therefore functions as a context-dependent shorthand rather than a single recognized concept across disciplines.

7. Cross-disciplinary significance

The different BLCE usages share almost no mathematical or conceptual common core beyond acronymic coincidence. The estimation-theoretic BLCE is a constrained minimum-variance linear unbiased estimator on an affine subspace (Lang et al., 2017). The computer-vision BLCE is an $y \in \mathbb{R}^{N_y}$41-trajectory inference module embedded in an end-to-end rendering-and-deblurring pipeline (Bui et al., 21 Apr 2025). The bandit BLCE is an elimination-based rare-update strategy with minimax-optimal regret up to polylogarithmic factors (Yu et al., 31 May 2026). The neutrino BLCE is a phenomenological mixing construction in which a CKM-like charged-lepton correction modifies a Bi-Large neutrino pattern (Roy et al., 2012). The astrophysical BLCE channel is an evolutionary interpretation of BLAPs based on post-common-envelope contraction and radiative levitation (Byrne et al., 2018).

A plausible implication is that references to “BLCE” without an expansion are intrinsically ambiguous in interdisciplinary settings. For arXiv-oriented readers, the safe interpretation is therefore bibliographic rather than lexical: BLCE should be resolved from the paper’s field, equations, and nearby terminology, not from the acronym alone.