Multiplier processes keep the observations fixed and randomize their influence through external weights. They provide a computational approximation to a Gaussian empirical-process limit and, unlike resampling observations, can be adapted to heteroskedastic scores and estimated influence functions.
Centered multiplier processes
Definition. Let \(\xi_1,\ldots,\xi_n\) be i.i.d., independent of the data, with \[ E\xi_i=0,\qquad E\xi_i^2=1. \] The conditionally centered multiplier process is \[ \mathbb G_n^\xi f = \frac1{\sqrt n}\sum_{i=1}^n \xi_i\{f(X_i)-P_nf\}, \qquad f\in\mathcal F. \]
Centering by \(P_nf\) ensures invariance under addition of constants and gives conditional covariance \[ E_\xi(\mathbb G_n^\xi f\,\mathbb G_n^\xi g) =P_n\{(f-P_nf)(g-P_ng)\}. \] This converges to the covariance of the \(P\)-Brownian bridge whenever the relevant products obey a uniform law of large numbers.
The multiplier central limit theorem
Definition. The Lorentz \(L_{2,1}\) norm of a random variable is \[ \|\xi\|_{2,1} = \int_0^\infty\sqrt{P(|\xi|>t)}\,\mathrm{d}t. \] A finite \(2+\delta\) moment implies \(\|\xi\|_{2,1}<\infty\), while a finite second moment alone need not.
Theorem (multiplier Donsker theorem). Let \(\mathcal F\) be a measurable \(P\)-Donsker class and let the multipliers satisfy \[ E\xi=0,\qquad E\xi^2=1,\qquad\|\xi\|_{2,1}<\infty. \] Then, conditionally on \(X_1,X_2,\ldots\), \[ \mathbb G_n^\xi\rightsquigarrow\mathbb G_P \qquad\text{in }\ell^\infty(\mathcal F), \] in probability, in the bounded-Lipschitz metric.
Proof outline. For fixed \(f_1,\ldots,f_k\), conditional Lindeberg–Feller applies to \[ n^{-1/2}\xi_i \bigl(f_1(X_i)-P_nf_1,\ldots,f_k(X_i)-P_nf_k\bigr)^\top. \] The conditional covariance converges by the law of large numbers. Truncate the multipliers at successively increasing levels. The bounded part is controlled by symmetrization and the asymptotic equicontinuity of the original Donsker class; the \(L_{2,1}\) condition makes the sum of truncation errors finite across dyadic levels. This proves conditional asymptotic equicontinuity. Finite-dimensional convergence plus tightness identifies \(\mathbb G_P\). \(\square\)
Conditional convergence
Definition. Let \(BL_1\) be the functions \(h:\ell^\infty(\mathcal F)\to[-1,1]\) with Lipschitz constant at most \(1\). Conditional weak convergence in probability means \[ \sup_{h\in BL_1} \left| E_\xi h(\mathbb G_n^\xi)-Eh(\mathbb G_P) \right| \xrightarrow{p}0, \] together with the corresponding asymptotic measurability condition when outer expectations are needed.
This formulation says that the random conditional law generated by the multipliers consistently estimates the fixed law of the Gaussian bridge. It is stronger than showing that the unconditional joint distribution has the right marginal limit.
Multiplier inequalities
Proposition. Let \(|\xi|_{(1)}\geq\cdots\geq|\xi|_{(n)}\) denote ordered multiplier magnitudes. For a class containing \(0\), a multiplier inequality bounds \[ E\left\| \frac1{\sqrt n}\sum_{i=1}^n\xi_i\{f(X_i)-Pf\} \right\|_{\mathcal F} \] by weighted partial-sample Rademacher suprema. In particular, if \[ E\left\| \frac1{\sqrt k}\sum_{i=1}^k\varepsilon_i \{f(X_i)-Pf\} \right\|_{\mathcal F} \leq C \quad\text{uniformly in }k, \] then the multiplier supremum is bounded by a constant multiple of \(C\|\xi\|_{2,1}\).
The Lorentz norm appears because the proof decomposes multipliers by order statistics and sums contributions over tail levels. This is the multiplier analogue of chaining over geometric resolutions.
Gaussian and Rademacher weights
Example (Gaussian multiplier). If \(\xi_i\sim N(0,1)\), then conditionally on the data \(\mathbb G_n^\xi\) is exactly Gaussian with empirical covariance kernel. Simulation requires only drawing a Gaussian vector and evaluating weighted sums. This is often the cleanest approximation for simultaneous confidence bands.
Example (Rademacher multiplier). If \(\xi_i\in\{-1,1\}\) with equal probabilities, the process is the same sign process used in symmetrization, after empirical centering. It is computationally stable and automatically bounded.
Example (wild weights). In heteroskedastic regression, an asymptotically linear statistic may have score contributions \(\hat\psi_i\) with observation-specific conditional variances. The wild multiplier statistic \[ T_n^\xi=\frac1{\sqrt n}\sum_{i=1}^n\xi_i\hat\psi_i \] preserves those magnitudes while randomizing signs. Mammen, Rademacher, or Gaussian weights are common choices; their first two moments determine the first-order limit, while higher moments affect finite-sample accuracy.
Estimated influence functions
Theorem. Suppose an estimator indexed by \(t\in T\) satisfies uniformly \[ \sqrt n\{\hat\theta_n(t)-\theta_0(t)\} =\mathbb G_n\psi_t+o_p(1) \quad\text{in }\ell^\infty(T), \] where \(\Psi=\{\psi_t:t\in T\}\) is \(P\)-Donsker. If estimated influence functions satisfy \[ \sup_{t\in T}\|\hat\psi_t-\psi_t\|_{P_n,2}=o_p(1) \] and belong with high probability to a class on which the multiplier process is asymptotically equicontinuous, then \[ t\mapsto\frac1{\sqrt n}\sum_{i=1}^n \xi_i\{\hat\psi_t(X_i)-P_n\hat\psi_t\} \] conditionally converges to the same Gaussian limit.
Quantiles of \(\sup_{t\in T}|\mathbb G_n^\xi\hat\psi_t|\) can then calibrate simultaneous confidence bands. The theorem also explains the two required checks: the original linear representation must be uniform, and replacing \(\psi_t\) by \(\hat\psi_t\) must be valid for the multiplier process, not only in pointwise \(L_2\).