Oriented Variational Forward (OVF)

• Oriented Variational Forward (OVF) is a family of methods that orients variation forward to drive optimization, ensuring monotonic descent and robust performance.
• OVF manifests differently across fields: in optimal control through virtual-time evolution, in variational inference via forward-KL divergence, and in NVCIM as negative feedback training.
• These frameworks offer practical gains including improved convergence, enhanced noise robustness, and calibrated posterior approximations in diverse computational settings.

Searching arXiv for papers on Oriented Variational Forward (OVF) and closely related formulations. Oriented Variational Forward (OVF) denotes a family of forward-oriented variational constructions appearing in distinct research contexts, rather than a single universally standardized method. In the literature provided here, the term has three technically different meanings. In optimal control, OVF corresponds to a forward evolution in a virtual variation time that drives trajectories toward extremals by monotone decrease of a functional, as formalized by the Variation Evolving Method (VEM) in “A Variation Evolving Method for Optimal Control” (Zhang et al., 2017). In likelihood-free variational inference, OVF refers to the forward-KL orientation underlying Forward Amortized Variational Inference (FAVI), where the variational objective uses $KL(p\|q)$ rather than $KL(q\|p)$ (Ambrogioni et al., 2018), with related forward-KL developments such as Transport Score Climbing (TSC) (Zhang et al., 2022). In non-volatile compute-in-memory deployment, OVF names a specific negative-feedback training mechanism for robustness against device variation in NVCIM accelerators (Qin et al., 2023, Qin et al., 17 Aug 2025). The common thread is an explicit orientation of variation or optimization in a forward direction, but the mathematical object being evolved differs substantially across these settings.

1. Terminological scope and research contexts

The term OVF is used explicitly in the NVCIM robustness literature as “Oriented Variational Forward,” where it denotes a training-time negative-feedback mechanism based on multiple noisy forward passes with increasing variation amplitudes (Qin et al., 2023, Qin et al., 17 Aug 2025). In the likelihood-free inference literature, the same label corresponds to the forward-KL orientation of FAVI: the “orientation” is $KL(p\|q)$, and the resulting objective is a joint-contrastive forward variational loss (Ambrogioni et al., 2018). In the optimal-control literature, the data identify OVF as a viewpoint on VEM: a “forward-oriented variational evolution (gradient flow) toward optimality,” implemented through evolution in a virtual time $\tau$ (Zhang et al., 2017).

These usages are not interchangeable. In VEM, the evolving object is a trajectory or optimal-control tuple such as $(x,u,\lambda,t_f)$; in FAVI and TSC, it is a variational distribution $q$; in NVCIM OVF, it is the training objective induced by combining a backbone output with negatively weighted noisy auxiliary forwards. A plausible implication is that OVF functions best as a cross-domain descriptor for methods that orient variation so as to enforce monotone descent or robustness, rather than as the name of one canonical algorithm.

2. OVF as virtual-time forward evolution in optimal control

In “A Variation Evolving Method for Optimal Control” (Zhang et al., 2017), the central construction is the introduction of a virtual “variation time” $\tau \in [0,\infty)$ that parameterizes a forward evolution of the candidate solution. For the calculus-of-variations problem,

$$J = \int_{t_0}^{t_f} F\big(y(t),\,\dot y(t),\,t\big)\,dt,$$

the method defines variation dynamics

$$\frac{\partial y}{\partial \tau}(t,\tau) = -K\Big(F_y - \frac{d}{dt}F_{\dot y}\Big),$$

with boundary evolution

$$\frac{\partial y}{\partial \tau}(t_0,\tau) = +K\,F_{\dot y}(t_0),\qquad \frac{\partial y}{\partial \tau}(t_f,\tau) = -K\,F_{\dot y}(t_f),$$

where $KL(q\|p)$0 is a positive diagonal gain matrix (Zhang et al., 2017). The functional $KL(q\|p)$1 is nonincreasing along $KL(q\|p)$2, and the equilibrium satisfies the Euler–Lagrange and boundary conditions.

For the optimal control problem with dynamics

$KL(q\|p)$3

and Bolza objective

$KL(q\|p)$4

the paper does not directly descend on an augmented Lagrangian. Instead, it constructs the convex residual functional

$KL(q\|p)$5

with residuals

$KL(q\|p)$6

and terminal residuals

$KL(q\|p)$7

where

$KL(q\|p)$8

is the Hamiltonian (Zhang et al., 2017). The evolution is then written as a functional gradient flow

$KL(q\|p)$9

for $KL(p\|q)$0 and $KL(p\|q)$1 (Zhang et al., 2017).

The paper explicitly relates this construction to an OVF viewpoint: VEM is described as “precisely a continuous-time gradient descent in a function space” and as conceptually identical to “oriented variational forward (OVF)” insofar as both orient variation to ensure monotonic decrease by forward integration in a virtual time (Zhang et al., 2017). In this setting, OVF is therefore a forward-evolution interpretation of function-space optimization for Euler–Lagrange and Pontryagin systems.

3. Stability, convergence, and numerical realization in the VEM/OVF viewpoint

The VEM formulation uses Lyapunov-style monotonicity arguments. For the calculus-of-variations case, taking $KL(p\|q)$2 yields

$KL(p\|q)$3

For the optimal-control case, taking $KL(p\|q)$4 gives

$KL(p\|q)$5

and, as stated in Theorem 2,

$KL(p\|q)$6

so the limit satisfies the PMP conditions and transversality conditions (Zhang et al., 2017). The paper emphasizes “global stability” in the sense that any initial guess respecting essential boundary conditions decreases $KL(p\|q)$7 or $KL(p\|q)$8 monotonically and converges to an extremal, though not necessarily a globally optimal one if the original OCP is nonconvex (Zhang et al., 2017).

Numerically, the method converts the original boundary-value problem into an initial-value problem in $KL(p\|q)$9. The “ZS equation” is interpreted as a PDE in $\tau$0, and in practice the implementation uses semi-discretization in physical time $\tau$1 and ODE integration in $\tau$2 (Zhang et al., 2017). The paper reports discretization on uniform grids, finite-difference approximations to $\tau$3 and $\tau$4, and integration with Matlab solvers: ode45 for the smooth calculus-of-variations example and ode15s for stiff OCP examples (Zhang et al., 2017).

The worked examples illustrate the forward-evolution mechanism. For a fixed-endpoint calculus-of-variations problem on $\tau$5, the numerical trajectory approaches the analytic solution while $\tau$6 decreases monotonically (Zhang et al., 2017). For a linear-quadratic-type OCP with fixed terminal time, numerical solutions converge to the analytic state, costate, and control, becoming indistinguishable from analytic by $\tau$7 s (Zhang et al., 2017). For the brachistochrone with free terminal time, the method converges to $\tau$8 s, close to the $\tau$9 s computed by GPOPS-II (Zhang et al., 2017).

This suggests that, in the optimal-control sense, OVF is best understood as an orientation principle for variational dynamics in infinite-dimensional trajectory spaces: the solution is not solved by direct shooting or collocation first, but approached asymptotically via forward evolution.

4. OVF as forward-KL orientation in variational inference

In the variational-inference literature, OVF refers to the forward orientation of the KL divergence. “Forward Amortized Inference for Likelihood-Free Variational Marginalization” (Ambrogioni et al., 2018) introduces Forward Amortized Variational Inference (FAVI), where the loss is obtained from the joint-contrastive forward KL between the model joint $(x,u,\lambda,t_f)$0 and a variational joint $(x,u,\lambda,t_f)$1. The forward KL is

$(x,u,\lambda,t_f)$2

and the joint-contrastive decomposition yields

$(x,u,\lambda,t_f)$3

Ignoring terms independent of $(x,u,\lambda,t_f)$4, the FAVI objective is

$(x,u,\lambda,t_f)$5

This is equivalent to minimizing the amortized forward posterior-contrastive loss

$(x,u,\lambda,t_f)$6

since the difference is again a term independent of $(x,u,\lambda,t_f)$7 (Ambrogioni et al., 2018).

The gradient estimator is

$(x,u,\lambda,t_f)$8

which the paper characterizes as unbiased and likelihood-free because it requires only samples $(x,u,\lambda,t_f)$9 and not evaluations of $q$0 or its derivatives (Ambrogioni et al., 2018). The paper further proves an exact-marginals theorem: for a fully factorized variational posterior

$q$1

the minimizer satisfies

$q$2

for all $q$3 and all $q$4 in the support of $q$5 (Ambrogioni et al., 2018). This marginal-optimality property is explicitly contrasted with reverse-KL mean-field inference.

The same forward-KL orientation appears in Transport Score Climbing (TSC) (Zhang et al., 2022), which optimizes

$q$6

with gradient

$q$7

TSC estimates this expectation using HMC samples from a warped posterior induced by an adaptive transport map $q$8, and the same map parameterizes the normalizing-flow family used for $q$9 (Zhang et al., 2022). The paper positions this as forward-KL variational inference that avoids uncertainty underestimation associated with reverse KL and leverages better transport maps to improve HMC, which in turn improves forward-KL optimization (Zhang et al., 2022).

In this literature, then, OVF is not a virtual-time PDE or negative-feedback training rule. It is the orientation of variational inference toward $\tau \in [0,\infty)$0, typically associated with mass-covering behavior, exact moment or marginal matching in specific families, and simulator-based or MCMC-based estimation of the forward-KL gradient (Ambrogioni et al., 2018, Zhang et al., 2022).

5. OVF as negative-feedback training for NVCIM accelerators

A third and distinct meaning appears in the NVCIM robustness literature. “Negative Feedback Training: A Novel Concept to Improve Robustness of NVCIM DNN Accelerators” (Qin et al., 2023) introduces Negative Feedback Training (NFT) and identifies OVF as one of two specific NFT instances, the other being Intermediate Representation Snapshot (IRS). “Special Session: Sustainable Deployment of Deep Neural Networks on Non-Volatile Compute-in-Memory Accelerators” (Qin et al., 17 Aug 2025) develops the same OVF mechanism in the context of sustainable deployment.

The motivation is the degradation of inference accuracy on non-volatile compute-in-memory substrates due to stochastic device-to-device and cycle-to-cycle programming variability, intrinsic device variations, and related physical non-idealities (Qin et al., 2023, Qin et al., 17 Aug 2025). The deployment mapping uses quantization and bit-slicing. For $\tau \in [0,\infty)$1 bits per weight, the quantized weight is

$\tau \in [0,\infty)$2

and with $\tau \in [0,\infty)$3 bits per device, the desired conductance of the $\tau \in [0,\infty)$4-th device is

$\tau \in [0,\infty)$5

Programming variation is modeled as

$\tau \in [0,\infty)$6

leading to a realized weight

$\tau \in [0,\infty)$7

(Qin et al., 17 Aug 2025). The experiments use $\tau \in [0,\infty)$8 bits per weight, $\tau \in [0,\infty)$9 bits per device, negative weights mapped to a separate array, and relative standard deviation bounded by $$J = \int_{t_0}^{t_f} F\big(y(t),\,\dot y(t),\,t\big)\,dt,$$0 (Qin et al., 17 Aug 2025).

OVF instantiates negative feedback by combining a backbone forward pass with multiple auxiliary noisy forwards of increasing noise amplitude. Let $$J = \int_{t_0}^{t_f} F\big(y(t),\,\dot y(t),\,t\big)\,dt,$$1 be the output computed with deployment-level variation, and let $$J = \int_{t_0}^{t_f} F\big(y(t),\,\dot y(t),\,t\big)\,dt,$$2 be oriented forwards under stronger variation. The total output is

$$J = \int_{t_0}^{t_f} F\big(y(t),\,\dot y(t),\,t\big)\,dt,$$3

where $$J = \int_{t_0}^{t_f} F\big(y(t),\,\dot y(t),\,t\big)\,dt,$$4 is the negative feedback strength, $$J = \int_{t_0}^{t_f} F\big(y(t),\,\dot y(t),\,t\big)\,dt,$$5 are decay factors, and $$J = \int_{t_0}^{t_f} F\big(y(t),\,\dot y(t),\,t\big)\,dt,$$6 scale the backbone and feedback contributions (Qin et al., 17 Aug 2025). The oriented forwards use

$$J = \int_{t_0}^{t_f} F\big(y(t),\,\dot y(t),\,t\big)\,dt,$$7

with the reported schedule $$J = \int_{t_0}^{t_f} F\big(y(t),\,\dot y(t),\,t\big)\,dt,$$8, increment $$J = \int_{t_0}^{t_f} F\big(y(t),\,\dot y(t),\,t\big)\,dt,$$9, and search range determined by start $$\frac{\partial y}{\partial \tau}(t,\tau) = -K\Big(F_y - \frac{d}{dt}F_{\dot y}\Big),$$0 and end $$\frac{\partial y}{\partial \tau}(t,\tau) = -K\Big(F_y - \frac{d}{dt}F_{\dot y}\Big),$$1 (Qin et al., 17 Aug 2025). The decay factors are

$$\frac{\partial y}{\partial \tau}(t,\tau) = -K\Big(F_y - \frac{d}{dt}F_{\dot y}\Big),$$2

and default contribution factors are

$$\frac{\partial y}{\partial \tau}(t,\tau) = -K\Big(F_y - \frac{d}{dt}F_{\dot y}\Big),$$3

(Qin et al., 17 Aug 2025).

The training objective is the expected task loss under the backbone variation and the oriented auxiliary variation distributions: $$\frac{\partial y}{\partial \tau}(t,\tau) = -K\Big(F_y - \frac{d}{dt}F_{\dot y}\Big),$$4 with Monte Carlo approximation by one backbone sample and $$\frac{\partial y}{\partial \tau}(t,\tau) = -K\Big(F_y - \frac{d}{dt}F_{\dot y}\Big),$$5 oriented samples per minibatch (Qin et al., 17 Aug 2025). Because all forwards share $$\frac{\partial y}{\partial \tau}(t,\tau) = -K\Big(F_y - \frac{d}{dt}F_{\dot y}\Big),$$6, the gradient contains a positive contribution from the backbone and negative contributions from the oriented forwards: $$\frac{\partial y}{\partial \tau}(t,\tau) = -K\Big(F_y - \frac{d}{dt}F_{\dot y}\Big),$$7 This is the sense in which the method is “negative feedback”: the stronger-noise auxiliary forwards act as subtractive constraints that orient training toward parameters less sensitive to device variation (Qin et al., 17 Aug 2025).

6. Empirical behavior, comparisons, and limitations in the NVCIM OVF literature

The NVCIM papers evaluate OVF on VGG-8 and ResNet-18 across MNIST, CIFAR-10, CIFAR-100, and Tiny ImageNet, using PyTorch, NVIDIA GPUs, Monte Carlo evaluation with 200 runs, and averages over at least five independent trainings; the reported results have 95% confidence interval $$\frac{\partial y}{\partial \tau}(t,\tau) = -K\Big(F_y - \frac{d}{dt}F_{\dot y}\Big),$$8 (Qin et al., 17 Aug 2025). Baselines are vanilla training without noise and Gaussian noise-injection training (Qin et al., 17 Aug 2025).

OVF is reported to improve Top-1 accuracy over Gaussian noise-injection training by up to 46.71% for VGG-8 on MNIST, 6.78% for VGG-8 on CIFAR-10, 5.35% for ResNet-18 on MNIST, 16.30% for ResNet-18 on CIFAR-10, 17.21% for ResNet-18 on CIFAR-100, and 3.22% for ResNet-18 on Tiny ImageNet (Qin et al., 17 Aug 2025). The earlier NFT paper reports the same OVF gains and additionally notes that IRS slightly exceeds OVF on some deeper-network cases, for example 20.24% on ResNet-18/CIFAR-10 and 20.10% on ResNet-18/CIFAR-100 (Qin et al., 2023). Epistemic uncertainty is measured through Expected Kullback–Leibler divergence (EKL), and OVF is reported to achieve higher accuracy while maintaining low EKL divergence, whereas Gaussian noise-injection training increases uncertainty even as it improves accuracy over vanilla (Qin et al., 17 Aug 2025). Under substantial device variation, OVF also shows improved convergence robustness; for VGG-8 on MNIST at $$\frac{\partial y}{\partial \tau}(t,\tau) = -K\Big(F_y - \frac{d}{dt}F_{\dot y}\Big),$$9, the reported count in 10 runs is 6 non-converging models for noise-injection versus 0 for OVF (Qin et al., 17 Aug 2025).

The papers interpret these results as evidence that structured, oriented negative feedback better matches the deterministic optimization process to the stochastic deployment environment than single-output Gaussian noise-injection alone (Qin et al., 2023, Qin et al., 17 Aug 2025). A plausible implication is that the “orientation” in OVF is not merely larger-noise augmentation, but a deliberate weighting scheme in which stronger perturbations impose stronger constraints through $$\frac{\partial y}{\partial \tau}(t_0,\tau) = +K\,F_{\dot y}(t_0),\qquad \frac{\partial y}{\partial \tau}(t_f,\tau) = -K\,F_{\dot y}(t_f),$$0 and the subtractive aggregation.

The limitations are also explicit. OVF adds $$\frac{\partial y}{\partial \tau}(t_0,\tau) = +K\,F_{\dot y}(t_0),\qquad \frac{\partial y}{\partial \tau}(t_f,\tau) = -K\,F_{\dot y}(t_f),$$1 extra forward passes per batch, so training time increases roughly by a factor of about $$\frac{\partial y}{\partial \tau}(t_0,\tau) = +K\,F_{\dot y}(t_0),\qquad \frac{\partial y}{\partial \tau}(t_f,\tau) = -K\,F_{\dot y}(t_f),$$2 relative to standard single-forward training (Qin et al., 17 Aug 2025). If $$\frac{\partial y}{\partial \tau}(t_0,\tau) = +K\,F_{\dot y}(t_0),\qquad \frac{\partial y}{\partial \tau}(t_f,\tau) = -K\,F_{\dot y}(t_f),$$3 is very small, OVF performs similarly to baselines while still incurring this training overhead (Qin et al., 17 Aug 2025). If the actual device variation deviates strongly from the assumed Gaussian model, re-tuning the orientation schedule and $$\frac{\partial y}{\partial \tau}(t_0,\tau) = +K\,F_{\dot y}(t_0),\qquad \frac{\partial y}{\partial \tau}(t_f,\tau) = -K\,F_{\dot y}(t_f),$$4, or enriching the noise model, may be necessary (Qin et al., 17 Aug 2025). The experiments focus on programming variability rather than explicitly simulating ADC/DAC noise, IR-drop, drift, or nonlinearity, although the papers state that OVF is compatible with richer forward-pass noise models (Qin et al., 17 Aug 2025).

7. Comparative interpretation across the three OVF meanings

Across these literatures, OVF always denotes an oriented use of variation, but the orientation principle is instantiated differently.

Context	Core object	Orientation mechanism
Optimal control / VEM (Zhang et al., 2017)	Trajectory or control-state-costate path	Forward evolution in virtual time $$\frac{\partial y}{\partial \tau}(t_0,\tau) = +K\,F_{\dot y}(t_0),\qquad \frac{\partial y}{\partial \tau}(t_f,\tau) = -K\,F_{\dot y}(t_f),$$5 decreasing $$\frac{\partial y}{\partial \tau}(t_0,\tau) = +K\,F_{\dot y}(t_0),\qquad \frac{\partial y}{\partial \tau}(t_f,\tau) = -K\,F_{\dot y}(t_f),$$6 or $$\frac{\partial y}{\partial \tau}(t_0,\tau) = +K\,F_{\dot y}(t_0),\qquad \frac{\partial y}{\partial \tau}(t_f,\tau) = -K\,F_{\dot y}(t_f),$$7
Variational inference / FAVI, TSC (Ambrogioni et al., 2018, Zhang et al., 2022)	Variational distribution $$\frac{\partial y}{\partial \tau}(t_0,\tau) = +K\,F_{\dot y}(t_0),\qquad \frac{\partial y}{\partial \tau}(t_f,\tau) = -K\,F_{\dot y}(t_f),$$8	Forward KL $$\frac{\partial y}{\partial \tau}(t_0,\tau) = +K\,F_{\dot y}(t_0),\qquad \frac{\partial y}{\partial \tau}(t_f,\tau) = -K\,F_{\dot y}(t_f),$$9 and gradients under $KL(q\|p)$00
NVCIM robustness (Qin et al., 2023, Qin et al., 17 Aug 2025)	Training objective for DNN deployment robustness	Negative-feedback combination of backbone and stronger-noise auxiliary forwards

In VEM, the forward direction is temporal in an auxiliary dimension. In FAVI and TSC, it is divergence-theoretic: the KL is oriented from target to approximation. In NVCIM OVF, it is robustness-oriented: auxiliary forwards are oriented toward larger device variation and subtracted during training.

These frameworks should therefore be distinguished carefully in technical writing. “OVF” in the NVCIM sense is not a synonym for forward-KL variational inference, and neither is identical to the optimal-control VEM viewpoint. The overlap lies in the meta-principle that variation is not sampled or perturbed arbitrarily, but oriented so as to induce monotone descent, posterior coverage, or robustness. This suggests a useful unifying interpretation: OVF is best regarded as a methodological motif in which a forward-oriented variational construction reshapes the optimization landscape to make desired equilibria—optimality conditions, calibrated posteriors, or noise-robust deployed models—more accessible.