Quadratic Logistic Models: A Unified Overview

• Quadratic logistic model is a family of models that embed logistic mechanisms with quadratic structures in responses, predictors, state dynamics, or objective surrogates.
• It spans various approaches including quadratic exponential regressions, sparse predictor interactions, network-dependent methods, and growth models with quadratic saturation.
• Researchers must clearly specify the locus of quadraticity—whether in outcomes, covariates, dynamics, or optimization—to avoid misinterpretation and ensure accurate estimation.

The expression quadratic logistic model does not denote a single canonical object. In recent arXiv usage it names several distinct constructions: a quadratic exponential family for multivariate binary responses with logistic full conditionals; logistic regression with explicit quadratic predictor terms or Ising-type network interactions; logistic growth laws containing a quadratic saturation term or a quadratic-in-time exponent; stochastic or fractional logistic systems with quadratic birth–death or competition mechanisms; and optimization procedures that replace the logistic log-likelihood by quadratic or piece-wise quadratic surrogates (Yong et al., 1 Oct 2025, Mukherjee et al., 2021, Chen, 27 Aug 2025, Hoadley, 2020, Anceschi et al., 2024). The common feature is the coexistence of a logistic mechanism with a quadratic structure, but the quadraticity may reside in the response distribution, the predictor space, the growth law, the state dynamics, or the numerical objective.

1. Terminological scope

A concise way to disambiguate the term is to identify where the quadratic structure enters.

Usage	Location of the quadratic structure	Representative source
Quadratic exponential logistic regression	Pairwise response products $y_i y_j$ in a multivariate binary distribution	(Yong et al., 1 Oct 2025)
Network-dependent logistic regression	Ising-type interaction $\mathbf{X}^\top \mathbf{A}\mathbf{X}$	(Mukherjee et al., 2021)
Sparse quadratic logistic regression	Predictor terms $x_i x_j$ in the linear predictor	(Shanmugam et al., 2017)
Quadratic logistic growth	Quadratic saturation $-\alpha X^2$ or exponent $-k t^2$	(Chen, 27 Aug 2025, Walker et al., 2024)
Quadratic/logistic dynamics	Conjugacy between $x \mapsto x^2+c$ and $x \mapsto r x(1-x)$	(Benyi et al., 14 Oct 2025)
Quadratic logistic optimization	Quadratic or piece-wise quadratic lower bounds to logistic loss	(Hoadley, 2020, Anceschi et al., 2024)

A recurrent misconception is that “quadratic logistic” always means logistic regression with squared covariates. In several of the cited usages, that interpretation is explicitly incorrect. In quadratic exponential logistic regression, the quadraticity is in the joint law of the responses rather than in the covariates; in SQP and MM formulations, it is in the optimization surrogate rather than in the model features (Yong et al., 1 Oct 2025, Hoadley, 2020).

2. Quadratic exponential logistic models for dependent binary responses

In multivariate binary analysis, the quadratic logistic model is often the quadratic exponential binary distribution (QEBD) and its regression counterpart. For an $m$-dimensional binary vector $\mathbf{Y}=(Y_1,\dots,Y_m)$, the joint model is

$$p(\mathbf{y};\theta)=\frac{1}{Z(\theta)}\exp\!\left\{\sum_{i=1}^m \theta_i y_i+\sum_{1\le i<j\le m}\theta_{ij}y_i y_j\right\},$$

with main effects $\mathbf{X}^\top \mathbf{A}\mathbf{X}$0, pairwise interactions $\mathbf{X}^\top \mathbf{A}\mathbf{X}$1, and an intractable normalizing constant $\mathbf{X}^\top \mathbf{A}\mathbf{X}$2. In graphical terms, $\mathbf{X}^\top \mathbf{A}\mathbf{X}$3 corresponds to absence of an edge. Each full conditional is logistic: $\mathbf{X}^\top \mathbf{A}\mathbf{X}$4 With covariates, a quadratic exponential logistic regression (QELR) conditional mean model takes the form

$\mathbf{X}^\top \mathbf{A}\mathbf{X}$5

with symmetry $\mathbf{X}^\top \mathbf{A}\mathbf{X}$6 required for compatibility of the conditionals (Yong et al., 1 Oct 2025).

The principal inferential difficulty is that direct maximum likelihood requires summation over $\mathbf{X}^\top \mathbf{A}\mathbf{X}$7 configurations. The pseudo-likelihood replaces the joint likelihood by the product of full conditionals and is computationally convenient, but naive GLM information-matrix standard errors are severely anti-conservative because the product of conditionals ignores the full joint dependence and reuses responses as both outcomes and regressors. The central result of the recent theory is that applying generalized estimating equations to the pseudo-likelihood with independence working correlation $\mathbf{X}^\top \mathbf{A}\mathbf{X}$8 yields consistent parameter estimates and valid sandwich standard errors, whereas dependent working correlations such as exchangeable or AR(1) may introduce non-ignorable bias because the estimating function no longer has mean zero unless the working covariance is diagonal. The estimating equations are

$\mathbf{X}^\top \mathbf{A}\mathbf{X}$9

and with $x_i x_j$0 the robust variance is

$x_i x_j$1

Simulation studies reported that GEE-IND was essentially unbiased and its standard errors closely matched empirical variability, whereas PL-as-GLM underestimated standard errors and non-diagonal working correlations could be biased or unstable. The same study reported substantial computational advantages for GEE-IND over MLE; for $x_i x_j$2, MLE required approximately $x_i x_j$3s at $x_i x_j$4, $x_i x_j$5s at $x_i x_j$6, and $x_i x_j$7s at $x_i x_j$8, while GEE-IND required approximately $x_i x_j$9s, $-\alpha X^2$0s, and $-\alpha X^2$1s, respectively. Applications included carcinogenic toxicity assays and constitutional court opinion writing data (Yong et al., 1 Oct 2025).

3. Predictor-space and network-dependent quadratic logistic regression

A different usage places the quadraticity in the predictor structure. In sparse quadratic logistic regression, the binary response satisfies

$-\alpha X^2$2

so the model contains both linear terms and explicit quadratic features $-\alpha X^2$3. The central algorithmic problem is support recovery when $-\alpha X^2$4 is large and a naive expansion to all $-\alpha X^2$5 quadratic features is infeasible. The proposed solution first recovers a weak support in $-\alpha X^2$6 time by correlation screening, then fits standard logistic regression on the reduced set of $-\alpha X^2$7 derived features. In the Boolean case this uses empirical correlations $-\alpha X^2$8, and in the non-binary finite-support case it uses a hashing-based nonlinear correlation test. The stated sample bound for the correlation screen is $-\alpha X^2$9 for polynomially small error probability, and the method was validated on both synthetic examples and the Dorothea dataset (Shanmugam et al., 2017).

In network-dependent logistic regression, the quadraticity instead represents peer effects over a known interaction graph. With spins $-k t^2$0, covariates $-k t^2$1, interaction matrix $-k t^2$2, scalar peer-effect parameter $-k t^2$3, and regression coefficients $-k t^2$4, the joint model is

$-k t^2$5

The corresponding conditional logit has linear predictor $-k t^2$6, where $-k t^2$7. Estimation proceeds by penalized maximum pseudo-likelihood,

$-k t^2$8

which avoids the partition function and targets sparsity only in $-k t^2$9. Under bounded network influence, non-degenerate design, bounded covariates and parameters, and a Frobenius-norm condition on $x \mapsto x^2+c$0, the joint estimator satisfies

$x \mapsto x^2+c$1

and under strengthened $x \mapsto x^2+c$2-boundedness one recovers the classical sparse rate

$x \mapsto x^2+c$3

This line therefore interprets the quadratic logistic model as an Ising-type regression with network dependence rather than as a classical polynomial-feature logit model (Mukherjee et al., 2021).

4. Deterministic growth laws and quadratic–logistic conjugacy

In growth modeling, a quadratic logistic model may denote a logistic curve with a quadratic temporal exponent. For fractal-dimension trajectories, the form

$x \mapsto x^2+c$4

is used in place of the ordinary logistic curve. Its growth-speed equation is

$x \mapsto x^2+c$5

so the instantaneous growth rate is $x \mapsto x^2+c$6, not a constant. The associated discrete regression

$x \mapsto x^2+c$7

permits ordinary least squares estimation of $x \mapsto x^2+c$8 and $x \mapsto x^2+c$9, hence $x \mapsto r x(1-x)$0 and $x \mapsto r x(1-x)$1, followed by a quasilinear regression of $x \mapsto r x(1-x)$2 on $x \mapsto r x(1-x)$3. In the Beijing multifractal application, the second-stage transformed regressions attained $x \mapsto r x(1-x)$4, $x \mapsto r x(1-x)$5, and $x \mapsto r x(1-x)$6 for $x \mapsto r x(1-x)$7, $x \mapsto r x(1-x)$8, and $x \mapsto r x(1-x)$9, respectively (Chen, 27 Aug 2025).

A distinct dynamical-systems usage arises from the affine conjugacy between the real quadratic family

$m$0

and the logistic family

$m$1

With $m$2, one has $m$3 provided

$m$4

This permits transfer of 3-cycle calculations between the quadratic map and the logistic map. In the quadratic variable $m$5, real 3-cycles exist iff $m$6; the cycle is born at $m$7 with multiplier $m$8, and loses stability at the unique real root $m$9 of

$\mathbf{Y}=(Y_1,\dots,Y_m)$0

where the multiplier is $\mathbf{Y}=(Y_1,\dots,Y_m)$1. Under the conjugacy, the biologically relevant logistic threshold is $\mathbf{Y}=(Y_1,\dots,Y_m)$2, and the stable period-3 window is

$\mathbf{Y}=(Y_1,\dots,Y_m)$3

This usage connects “quadratic” and “logistic” through conjugate one-dimensional maps rather than through regression or population growth (Benyi et al., 14 Oct 2025).

5. Stochastic, structured, fractional, and biomedical extensions

Several papers generalize logistic dynamics by introducing quadratic structure into the state evolution itself. One stochastic route is the logistic branching process with quadratic competition: in state $\mathbf{Y}=(Y_1,\dots,Y_m)$4, births occur at rate $\mathbf{Y}=(Y_1,\dots,Y_m)$5, natural deaths at rate $\mathbf{Y}=(Y_1,\dots,Y_m)$6, and competition deaths at rate $\mathbf{Y}=(Y_1,\dots,Y_m)$7. Its deterministic mean-field analogue is

$\mathbf{Y}=(Y_1,\dots,Y_m)$8

but the stochastic process goes extinct almost surely in finite time even when $\mathbf{Y}=(Y_1,\dots,Y_m)$9. Conditioning on survival produces explicit finite-time Doob transforms and a time-homogeneous Q-process, and the Yaglom distribution is characterized through a second-order ODE for its probability generating function (Etheridge et al., 2013).

A second route uses quadratic birth–death rates in discretely structured or multi-sigmoidal systems. In one formulation,

$$p(\mathbf{y};\theta)=\frac{1}{Z(\theta)}\exp\!\left\{\sum_{i=1}^m \theta_i y_i+\sum_{1\le i<j\le m}\theta_{ij}y_i y_j\right\},$$0

If $$p(\mathbf{y};\theta)=\frac{1}{Z(\theta)}\exp\!\left\{\sum_{i=1}^m \theta_i y_i+\sum_{1\le i<j\le m}\theta_{ij}y_i y_j\right\},$$1, $$p(\mathbf{y};\theta)=\frac{1}{Z(\theta)}\exp\!\left\{\sum_{i=1}^m \theta_i y_i+\sum_{1\le i<j\le m}\theta_{ij}y_i y_j\right\},$$2, and $$p(\mathbf{y};\theta)=\frac{1}{Z(\theta)}\exp\!\left\{\sum_{i=1}^m \theta_i y_i+\sum_{1\le i<j\le m}\theta_{ij}y_i y_j\right\},$$3, then the mean satisfies a multi-sigmoidal logistic law

$$p(\mathbf{y};\theta)=\frac{1}{Z(\theta)}\exp\!\left\{\sum_{i=1}^m \theta_i y_i+\sum_{1\le i<j\le m}\theta_{ij}y_i y_j\right\},$$4

Related structured-population analysis shows that moment closure is exact only under stringent polynomial constraints. Writing the moments as $$p(\mathbf{y};\theta)=\frac{1}{Z(\theta)}\exp\!\left\{\sum_{i=1}^m \theta_i y_i+\sum_{1\le i<j\le m}\theta_{ij}y_i y_j\right\},$$5, exact finite closure requires

$$p(\mathbf{y};\theta)=\frac{1}{Z(\theta)}\exp\!\left\{\sum_{i=1}^m \theta_i y_i+\sum_{1\le i<j\le m}\theta_{ij}y_i y_j\right\},$$6

together with coefficient identities such as

$$p(\mathbf{y};\theta)=\frac{1}{Z(\theta)}\exp\!\left\{\sum_{i=1}^m \theta_i y_i+\sum_{1\le i<j\le m}\theta_{ij}y_i y_j\right\},$$7

If $$p(\mathbf{y};\theta)=\frac{1}{Z(\theta)}\exp\!\left\{\sum_{i=1}^m \theta_i y_i+\sum_{1\le i<j\le m}\theta_{ij}y_i y_j\right\},$$8, exact closure is impossible at any finite order $$p(\mathbf{y};\theta)=\frac{1}{Z(\theta)}\exp\!\left\{\sum_{i=1}^m \theta_i y_i+\sum_{1\le i<j\le m}\theta_{ij}y_i y_j\right\},$$9 (Crescenzo et al., 2024, Walker et al., 2024).

Fractional and biomedical variants retain the quadratic logistic term while altering the differential operator or the ambient state space. In the ABC-fractional setting, the quadratic logistic equation is

$\mathbf{X}^\top \mathbf{A}\mathbf{X}$00

with equilibria $\mathbf{X}^\top \mathbf{A}\mathbf{X}$01 and $\mathbf{X}^\top \mathbf{A}\mathbf{X}$02; the reported linearized analysis yields instability of $\mathbf{X}^\top \mathbf{A}\mathbf{X}$03 and asymptotic stability of $\mathbf{X}^\top \mathbf{A}\mathbf{X}$04. A carrying-capacity variant,

$\mathbf{X}^\top \mathbf{A}\mathbf{X}$05

shows the same equilibrium pattern numerically, with convergence speed modulated by the kernel parameters (Abdeljawad et al., 2019). In an HIV model with two cell-compartment ODEs and a two-dimensional PDE for virions, the T-cell equation contains

$\mathbf{X}^\top \mathbf{A}\mathbf{X}$06

The uninfected equilibrium is governed by

$\mathbf{X}^\top \mathbf{A}\mathbf{X}$07

and the reported stability criterion is $\mathbf{X}^\top \mathbf{A}\mathbf{X}$08 for asymptotic stability of the disease-free state. Viral diffusion in two dimensions does not alter that threshold and does not generate diffusion-driven pattern instability; Hopf bifurcations, when present, produce spatially homogeneous periodic solutions (Brauner et al., 2012).

6. Quadratic surrogates in logistic optimization

In computational statistics, the phrase may refer not to the model class but to the optimization architecture used to fit logistic models. In score-engineered logistic regression, the negative log-likelihood

$\mathbf{X}^\top \mathbf{A}\mathbf{X}$09

is locally replaced by its second-order Taylor expansion at $\mathbf{X}^\top \mathbf{A}\mathbf{X}$10,

$\mathbf{X}^\top \mathbf{A}\mathbf{X}$11

yielding a convex quadratic program at each iteration. This is the basis of a sequential quadratic programming framework for constrained score engineering with monotonicity, convexity, centering, and box constraints. In this usage, the quadraticity lies in the surrogate loss, not in the features; the paper explicitly distinguishes this meaning from logistic regression with quadratic predictor terms (Hoadley, 2020).

A related MM literature studies tangent lower bounds for the logistic log-likelihood via the one-dimensional concave function

$\mathbf{X}^\top \mathbf{A}\mathbf{X}$12

Classical quadratic minorizers include the Böhning bound and the Jaakkola–Jordan/Pólya–Gamma bound, but a newer construction introduces the piece-wise quadratic lower bound

$\mathbf{X}^\top \mathbf{A}\mathbf{X}$13

with

$\mathbf{X}^\top \mathbf{A}\mathbf{X}$14

The stated domination result is

$\mathbf{X}^\top \mathbf{A}\mathbf{X}$15

so the piece-wise quadratic minorizer uniformly improves any tangent quadratic minorizer while remaining tractable. In ridge logistic regression, this sharper bound reduced iterations markedly relative to Böhning and Pólya–Gamma schemes on the Leukemia and PRISM examples, and it admits a direct reformulation as a generalized lasso problem (Anceschi et al., 2024).

The optimization-based usage therefore stands at the opposite end of the spectrum from QEBD or growth-law interpretations. Here the logistic model may be entirely standard; “quadratic” refers to the algorithmic local model used for estimation.

7. Conceptual unification

Across these literatures, the term can be organized by asking which of four objects is made quadratic.

First, the joint response law may be quadratic, as in QEBD and QELR, where pairwise response products $\mathbf{X}^\top \mathbf{A}\mathbf{X}$16 generate Ising-type dependence and logistic full conditionals. Second, the predictor representation may be quadratic, either through explicit terms $\mathbf{X}^\top \mathbf{A}\mathbf{X}$17 in sparse quadratic logistic regression or through network interactions $\mathbf{X}^\top \mathbf{A}\mathbf{X}$18 in Ising-regression hybrids. Third, the state dynamics may be quadratic, through saturation terms such as $\mathbf{X}^\top \mathbf{A}\mathbf{X}$19, competition terms $\mathbf{X}^\top \mathbf{A}\mathbf{X}$20, quadratic birth–death rates, or quadratic logistic PDE components. Fourth, the objective function approximation may be quadratic, through SQP or MM surrogates for logistic log-likelihoods (Yong et al., 1 Oct 2025, Shanmugam et al., 2017, Etheridge et al., 2013, Hoadley, 2020).

This suggests that “quadratic logistic model” is best treated as a family resemblance term rather than a uniquely defined model class. The phrase is precise only after the locus of the quadratic structure is specified. In regression, the crucial distinction is whether the quadraticity concerns covariates, outcomes, or numerical approximation; in dynamics, whether it concerns state saturation, time dependence, or conjugacy with a quadratic map; and in stochastic modeling, whether it concerns interaction rates, conditioning, or memory kernels. Without that specification, the term is intrinsically ambiguous in current arXiv practice (Mukherjee et al., 2021, Chen, 27 Aug 2025, Benyi et al., 14 Oct 2025, Anceschi et al., 2024).