Pull-to-Center Biases

• Pull-to-center biases are systematic shifts in estimates or behaviors toward a central, prototypical reference point, observable in fields such as perception and computational sampling.
• They emerge from mechanisms like perceptual summary operations, algorithmic irreversibility, and context-dependent optimization, leading to non-uniform distributions.
• Mitigation strategies include experimental design adjustments and algorithmic corrections to counteract central clustering in perceptual reports and simulation outputs.

Pull-to-center biases are a class of systematic distortions in which estimates, behaviors, or sampled configurations are drawn toward a central, prototypical, or otherwise salient reference point within a defined space. Manifesting across domains as diverse as computational sampling, human perception, social cognition, and astronomical data analysis, these biases are generically characterized by a non-uniform probability or response distribution with elevated density near the geometrical or contextual center, or by errors that interpolate between multiple attractor points. Pull-to-center biases may emerge from mechanisms such as algorithmic irreversibility, perceptual summary operations, transient attention maps, or context-dependent statistical optimization.

1. Mathematical and Mechanistic Definitions

The formalization of pull-to-center biases is context dependent but typically involves a shift of estimates (gaze, centroids, memory, or sampling) toward a central location. In social cognition, Bavaud introduces the in-focus ("pull-to-center") and out-focus ("push-away") points to describe how context alters perceived group tendencies by minimizing or maximizing the relative inertia of group $g$ with respect to an overall sample $f$:

• For $x_i\in \mathbb{R}^p$, $f_i, g_i$ as statistical weights,
• Squared Euclidean distance $D_{ij} = \sum_{k=1}^p (x_{ik} - x_{jk})^2$,
• Inertia about $a$: $\Delta_f^a = \sum_{i=1}^n f_i D_{ia}$,
• Relative inertia (dispersion): $\delta(a) = \Delta_g^a / \Delta_f^a$,
• The in-focus $a_{-}$ is the unique minimizer $\operatorname{argmin}_a \delta(a)$, always shifted toward the global centroid $\bar x_f$: this geometric pull is the analytic core of stereotypic polarization (Bavaud, 2010).

In empirical perception, as in visual graph interpretation, the perceptual pull effect quantifies reporting bias as an interpolation:

$$\text{reported} = (1-\lambda)\cdot\mu_{\text{target}} + \lambda\cdot\mu_{\text{nontarget}}, \quad \lambda > 0,$$

where the recalled mean of a target series is displaced toward the mean of a non-target series (Xiong et al., 2019).

In sampling algorithms (e.g., lattice polymers), an irreversibility flaw in move proposals can cause the stationary distribution for measures such as end-to-end distance to skew toward more compact, centralized conformations—a kinetic "pull-to-center" (Györffy et al., 2012).

2. Behavioral and Perceptual Manifestations

Pull-to-center biases are prominent in human perception and action:

• Central Fixation Bias (CFB): When viewing images, human gaze demonstrates a transient, strong preference for the image center immediately following sudden onset. Rothkegel et al. find that the spatial distribution of initial fixations is markedly clustered at the geometric midpoint, regardless of scene content, viewing task, or start position (Rothkegel et al., 2016). This bias is dictated by a default, time-decaying activation centered on the new stimulus, superimposed on salience and task-driven guidance.
• Visual Average Estimation: When reporting the mean position of graphical elements following a brief exposure, observers systematically underestimate means for lines (MD = –4.49 px), overestimate for bars (MD = +4.19 px), and—critically—display "perceptual pull": their reported mean for one series is biased toward the concurrently displayed non-target, consistent with a linear interpolation model (Xiong et al., 2019).
• Categorical Position and Memory Anchoring: Memory for object positions is biased toward spatial category centers, mirroring the clustering in fixation and average reporting (Xiong et al., 2019).

3. Algorithmic and Statistical Sources

Pull-to-center biases can be generated by design flaws or optimization in statistical or algorithmic procedures:

• Monte Carlo Sampling (Lattice Polymers): The widely used "pull move" set in HP lattice protein models was considered reversible but, as Györffy et al. demonstrated, certain end pull moves create configurations (hooked ends) with no corresponding reverse. This violation of detailed balance drives systematic oversampling of compact (centralized) states, biasing the density of states $g(E)$ at low energies by orders of magnitude and reducing measures such as radius of gyration $R_g$ (Györffy et al., 2012).
• Resolution: Excluding all pull moves generating hooked ends restores reversibility and eliminates the bias, rigorously ensuring correct Boltzmann sampling (Györffy et al., 2012).
• Statistical Group Judgments: In stereotype formation, minimizing relative inertia generates in-focus points that systematically shift group summaries toward the sample centroid. No bias occurs for dispersion-free singleton groups; the pull magnitude depends analytically on group dispersion and separation between group and overall centroids (Bavaud, 2010).

4. Models and Quantitative Characterization

Model-based approaches have been used to formalize and quantify pull-to-center biases across modalities.

Human Fixation Modeling

Model Type	Pull-to-Center Accounted For?	Comments
Density Sampling	No	Stationary, no initial central pull
Gaussian Window	No	Insufficient onset dip
Center-Bias Model	Weak	Sluggish, fails to match rapid shift
Original SceneWalk	No	Matches later exploration only
SceneWalk StartMap	Yes	Default central activation, early decay

Only SceneWalk StartMap, which initializes the attention map with a central elliptical Gaussian, reproduces both the temporal onset and decay of CFB, fitting human data for early fixations (Rothkegel et al., 2016).

Visual Estimation Bias

Biases are assessed via mean signed error (MSE), linear mixed-effects models, and partial eta squared ($\eta^2_\text{partial}$). For superimposed series in graphs, the interaction effect (pull) shows up as a significant shift in MSE toward the non-target, with modest but robust effect sizes ($\eta^2_\text{partial} \sim 0.003{-}0.014$) (Xiong et al., 2019).

Stereotype Model

Closed-form expressions for in-focus and out-focus points, parameterized on the line between group and sample centroid, yield quantifiable shifts:

$$a_{-} = \bar x_g + (1+\epsilon_{-})(\bar x_f-\bar x_g), \quad 1+\epsilon_{-} \le 0,$$

with the magnitude and direction determined by relative group and total inertia and their separation (Bavaud, 2010).

5. Domain-Specific Examples

Weak Lensing Mass Estimates

Pull-to-center bias manifests as mass overestimation when the observationally identified cluster center (from X-ray/SZE) does not coincide with the "true" center. Sommer et al. show that the miscentering distribution is typically anisotropic, leading to directional mass bias:

• For SZE-derived centers as reference, bias $⟨b⟩ \approx 1.02$, i.e., +2% overestimate relative to the potential minimum.
• For randomized miscentering (isotropic), bias $⟨b⟩ \approx 0.97$, i.e., –3% underestimate.
• Net over-correction reaches $\tau ≈ +6\%$, modeled as $\tau(σ) ≈ A (σ/r_s)^{α}$, with $A ≈ 0.1$, $α ≈ 1.2$; importantly, adopting the 3D center of mass as reference reduces both mean offset and anisotropy, rendering $τ ≲ 1\%$ (Sommer et al., 2024).

Visual Graphs and Scene Viewing

Perceptual pull alters average position reporting in data visualizations, with practical consequences for chart design and information recall. The presence of multiple data series exacerbates these biases, indicating involuntary integration during summary judgments (Xiong et al., 2019).

6. Mitigation and Experimental/Algorithmic Controls

Multiple strategies have been empirically validated or prescribed to reduce or account for pull-to-center biases:

• Experimental Paradigms (Scene Viewing):
  • Introduce short enforced fixation delays (e.g., ≥125 ms) after stimulus onset, which attenuates central fixation bias to near-baseline levels.
  • Vary start position uniformly or peripherally, avoiding always starting at center.
  • Track and analyze saccade latencies, explicitly modeling or excluding rapid-onset fixations (Rothkegel et al., 2016).
• Algorithmic Correction (Lattice Polymers):
  • Rigorously exclude non-reversible moves (e.g., hooked-end pulls), ensuring all transition steps possess a well-defined inverse (Györffy et al., 2012).
• Visualization Recommendations:
  • Employ bar charts for better average recall performance (smaller, more stable biases than lines).
  • Avoid concurrent multi-series overlays unless necessary; prefer separate axes or augmented with explicit markers for means.
  • Increase inter-series spacing or encoding orthogonality to reduce inter-series pull (Xiong et al., 2019).
• Statistical Stereotype Modeling:
  • Recognize that group typicality as judged via in-focus points is inherently context-dependent and geometrically pulled toward sample centroids, not a neutral property of the group alone (Bavaud, 2010).

7. Implications and Broader Significance

Pull-to-center biases constitute a substantial and domain-general source of error, distorting scientific measurement, human inference, algorithmic sampling, and social judgment. While their mechanistic origins vary—from reversible Markov transitions in sampling, to dynamic attention-state initialization in gaze, to utility-optimizing statistical inference in social perception—their observable impact is reliably manifested as a shift toward salient, central, or contextual reference positions. Proper experimental, analytic, and computational recognition of these effects is critical for the accuracy and interpretability of results in both the physical, cognitive, and social sciences.

References:

• "Biased Average Position Estimates in Line and Bar Graphs: Underestimation, Overestimation, and Perceptual Pull" (Xiong et al., 2019)
• "Directional miscentering dependence in weak lensing mass bias" (Sommer et al., 2024)
• "Stereotype bias: a simple formal model" (Bavaud, 2010)
• "The temporal evolution of the central fixation bias in scene viewing" (Rothkegel et al., 2016)
• "Pull moves for rectangular lattice polymer models are not fully reversible" (Györffy et al., 2012)