Newfluence: High-D Influence Correction

• Newfluence is a high-dimensional correction that refines classical influence functions by accounting for non-negligible shrinkage factors in modern models.
• It employs a one-step Newton update with the Woodbury formula to closely approximate leave-one-out retraining outcomes without excessive computation.
• Empirical tests in logistic regression show that Newfluence achieves near-perfect rank correlations in high-dimensional settings, particularly with moderate regularization.

Newfluence is a high-dimensional correction to classical influence functions for estimating how much an individual training example affects a model’s parameters, predictions, or test loss. It was introduced to address the mismatch between the low-dimensional asymptotics underlying classical influence functions and modern settings in which the parameter dimension $p$ is not much smaller than the sample size $n$, or even of the same order. In the proportional asymptotic regime $n,p\to\infty$ with $n/p\to\gamma_0$, the standard influence-function approximation is systematically biased by a non-negligible, data-point-specific shrinkage factor $1-H_{ii}$, whereas Newfluence corrects this while retaining similar computational efficiency (Zou et al., 16 Jul 2025).

1. Problem domain and interpretive role

Newfluence belongs to the literature on data attribution, model interpretability, and debugging. The underlying question is how to estimate the effect of deleting one training point $z_i$ on a downstream quantity, typically the learned parameter vector, the prediction at a test point, or the test loss on a point $z_0$. In the leave-one-out formulation, if $\hat\beta$ is trained on all data and $\hat\beta_{/i}$ is trained after removing $z_i$, the target quantity is the true influence

$n$0

Exact computation requires retraining $n$1 times, which is often infeasible, so classical influence functions replace it with a first-order approximation based on gradients and an inverse Hessian (Zou et al., 16 Jul 2025).

The motivation for Newfluence is that the classical approximation is justified by a regime in which $n$2. In modern ML and AI models, however, high-dimensionality makes leave-one-out perturbations non-infinitesimal in the relevant asymptotic sense. Newfluence therefore targets the same interpretive use cases as classical influence functions—identifying influential training examples, debugging mispredictions, model interpretability through data attribution, and data valuation or auditing—but reanalyzes the approximation under proportional asymptotics rather than under low-dimensional heuristics (Zou et al., 16 Jul 2025).

2. High-dimensional failure of classical influence functions

In the regularized empirical risk minimization setting studied by Newfluence, the estimator is

$n$3

and the leave-one-out estimator is

$n$4

The classical influence-function approximation uses the Hessian

$n$5

and yields

$n$6

For generalized linear models,

$n$7

The high-dimensional obstruction is the leverage-like quantity

$n$8

Under the proportional asymptotic regime,

$n$9

so $n,p\to\infty$0 is not approximately $n,p\to\infty$1. The central bias statement is

$n,p\to\infty$2

This implies that the classical estimate is systematically shrunken by a factor that remains order one in high dimension. The paper further notes that for many common losses and regularizers,

$n,p\to\infty$3

so the bias can be substantial when regularization is only moderate, and becomes small only in more effectively low-dimensional regimes induced by very strong regularization (Zou et al., 16 Jul 2025).

3. Formal setting and assumptions

The theory is developed for generalized linear models fit by regularized empirical risk minimization. Data are

$n,p\to\infty$4

with $n,p\to\infty$5, $n,p\to\infty$6, and

$n,p\to\infty$7

The loss shorthand is

$n,p\to\infty$8

with

$n,p\to\infty$9

The assumptions include a separable regularizer

$n/p\to\gamma_0$0

twice differentiability of $n/p\to\gamma_0$1 and $n/p\to\gamma_0$2, convexity with $n/p\to\gamma_0$3 $n/p\to\gamma_0$4-strongly convex, polynomial-growth and Lipschitz conditions, Gaussian covariates

$n/p\to\gamma_0$5

and response tails

$n/p\to\gamma_0$6

with $n/p\to\gamma_0$7, $n/p\to\gamma_0$8. The asymptotic regime is

$n/p\to\gamma_0$9

This setting is narrower than “modern AI models” in the colloquial sense. The rigorous guarantees currently cover generalized linear models with strongly convex, twice-differentiable objectives. The paper states explicitly that the present theory does not yet cover non-convex deep neural networks, non-smooth objectives, or sequential and reinforcement-learning settings (Zou et al., 16 Jul 2025).

4. Construction of Newfluence

Newfluence replaces the classical infinitesimal perturbation argument with a one-step Newton approximation to the leave-one-out estimator. The leave-one-out Hessian is

$1-H_{ii}$0

Since $1-H_{ii}$1 and $1-H_{ii}$2,

$1-H_{ii}$3

A one-step Newton update from $1-H_{ii}$4 gives

$1-H_{ii}$5

Using

$1-H_{ii}$6

the Woodbury formula yields

$1-H_{ii}$7

This is the corrected approximation to $1-H_{ii}$8. Newfluence then evaluates the loss at this corrected parameter, rather than linearizing the test loss a second time: $1-H_{ii}$9

A linearized correction also appears in the paper,

$z_i$0

and it is noted that this corrected version is consistent for the true influence. Newfluence nonetheless uses the nonlinear loss evaluation

$z_i$1

because this avoids the extra first-order Taylor approximation in the test loss. The expensive object remains the inverse-Hessian action $z_i$2, so the paper’s computational claim is that Newfluence preserves essentially the same order of effort as classical influence functions while avoiding $z_i$3 leave-one-out retrainings (Zou et al., 16 Jul 2025).

5. Accuracy guarantees and empirical behavior

The main theorem states three asymptotic facts. First, Newfluence is highly accurate: $z_i$4 Second, the true influence itself is larger: $z_i$5 Third, the classical influence function misses the multiplicative factor discussed above: $z_i$6 Since the error of the classical estimator is of the same order as the signal in high dimension, the distortion is not asymptotically negligible (Zou et al., 16 Jul 2025).

Empirically, the paper evaluates $z_i$7-regularized logistic regression in synthetic high-dimensional binary classification with

$z_i$8

$z_i$9

$z_0$0

and $z_0$1. For $z_0$2 unseen test examples, training-point rankings induced by Newfluence and by the classical influence function are compared against exact leave-one-out rankings using Kendall’s $z_0$3. In the weakly regularized high-dimensional regime $z_0$4, with

$z_0$5

Newfluence achieves

$z_0$6

while the classical influence function remains around

$z_0$7

In the strongly regularized regime $z_0$8, with

$z_0$9

both methods achieve

$\hat\beta$0

The empirical interpretation is the same as the theorem’s: high-dimensional bias is visible when effective complexity remains high, and vanishes when regularization makes the problem effectively low-dimensional (Zou et al., 16 Jul 2025).

6. Relation to adjacent influence frameworks and broader significance

Newfluence is a model-interpretability method for training-data attribution, not a social-network influence metric. This distinction matters because “influence” has parallel meanings across research areas. In social-media analysis, for example, influence may be defined causally as the average per-vertex increase in narrative tweets attributable to a source’s participation in a networked diffusion process (Smith et al., 2018), or dynamically as algorithmically mediated popularity trajectories in creator ecosystems (Galante et al., 4 Jul 2025). Newfluence instead studies the effect of removing one training example from a supervised learning problem (Zou et al., 16 Jul 2025).

Within ML influence estimation, Newfluence is also distinct from work that centers training randomness rather than high-dimensional bias. The $\hat\beta$1-INE framework introduces $\hat\beta$2-influence as a hypothesis-testing notion based on distinguishability between training on $\hat\beta$3 and training on $\hat\beta$4, with $\hat\beta$5-influence as a Gaussian specialization under composition (Panda et al., 12 Oct 2025). Newfluence does not redefine influence as a trade-off function; it retains the leave-one-out target but corrects the asymptotic approximation used to estimate it in the $\hat\beta$6 regime (Zou et al., 16 Jul 2025). A plausible implication is that the two lines address orthogonal failure modes: one concerns high-dimensional deterministic bias in the approximation itself, the other concerns instability under randomized training.

The broader significance of Newfluence is twofold. Methodologically, it shows that low-dimensional robust-statistics heuristics cannot simply be ported unchanged into proportional-asymptotic regimes. Conceptually, it suggests that attribution methods built from perturbative leave-one-out reasoning—an observation the paper explicitly extends to Shapley values as a possible future analytical target—may require their own high-dimensional reanalysis (Zou et al., 16 Jul 2025). The current scope remains limited to generalized linear models with smooth, strongly convex objectives, but within that scope Newfluence provides a precise correction: $\hat\beta$7 and establishes that the correction removes the order-one shrinkage missed by the classical influence function in high dimensions.