Lifting Technique: Methods & Applications

• Lifting technique is a design pattern that reformulates problems in higher-dimensional spaces to expose linearity, convexity, and tractable structure.
• It underpins advanced methods in neural approximation, geometric pose estimation, and optimization by enabling convexity and efficient computation.
• Practical implementations range from spline-based deep learning layers to robotic feasibility studies and integer programming cut-strengthening.

“Lifting technique” is a field-dependent term applied to constructions that move a problem from its original representation into a higher-dimensional, function-space, block-structured, or otherwise augmented formulation. In the cited literature, this includes spline-generating neural layers and lifted losses in deep learning (Ochs et al., 2018), 2D-to-3D pose reconstruction in computer vision (Hardy et al., 2023, Zhou et al., 2023), coefficient generation for strengthened inequalities in integer programming and bilinear optimization (Prasad et al., 2024, Gu et al., 2021, Wang et al., 26 Feb 2026), trajectory-space and observable-space reformulations in control and system identification (Yamamoto et al., 16 Sep 2025, Gerdpratoom et al., 10 Jan 2025, Mauroy et al., 2016, Mauroy et al., 2017, Bolli et al., 6 May 2026, Katayama et al., 2021), and recursive constructions in algebraic statistics, coded caching, game theory, scientific software verification, and graph-model compression (Rauh et al., 2014, Aravind et al., 2023, Kozachinskiy, 2021, Zhang et al., 15 Jan 2025, Sourek et al., 2020). This suggests that “lifting” is not a single method but a recurrent design pattern whose purpose is to expose linearity, convexity, compatibility, symmetry, or tractable structure.

1. Neural approximation and lifted losses

In neural approximation, lifting was formalized as a nonlinear transfer function that increases the dimensionality of the input. For a knot sequence $t^1<\dots<t^L$, scalar lifting maps $x\in[t^\ell,t^{\ell+1}]$ to

$$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$

with inverse

$$\psi(z)=\sum_{i=1}^L z_i t^i, \qquad \psi(\phi(x))=x.$$

A vector-valued analogue uses a triangulation of $\Omega\subset\mathbb R^d$ and barycentric coordinates; on each simplex, $\phi(x)$ has exactly $d+1$ nonzero entries (Ochs et al., 2018).

The central structural result is that a one-layer network

$f_\theta(x)=\langle \theta,\phi(x)\rangle$

or, for vector output, $f_\theta(x)=\theta\,\phi(x)$, is a continuous piecewise-linear function with breakpoints at the knots, and conversely any piecewise-linear function on that knot set or triangulation can be written in this form. The same framework yields a convex fitting problem in $\theta$ whenever the loss is convex. The paper also lifts the output loss: with output-lifting $x\in[t^\ell,t^{\ell+1}]$0 and simplex constraint $x\in[t^\ell,t^{\ell+1}]$1, the lifted loss

$x\in[t^\ell,t^{\ell+1}]$2

produces a convex linear program in $x\in[t^\ell,t^{\ell+1}]$3 for a class of non-convex or flat losses (Ochs et al., 2018).

The implementation is correspondingly simple. The forward pass identifies the interval $x\in[t^\ell,t^{\ell+1}]$4, writes the lifted vector with only two nonzero entries, and the backward pass differentiates through the linear interpolation weight. Hyperparameter guidance in the same work states that small $x\in[t^\ell,t^{\ell+1}]$5 often suffices in deep nets, while larger $x\in[t^\ell,t^{\ell+1}]$6 gives finer piecewise-linear approximation at cost of memory; knot placement may be uniform over the expected range or data-driven by quantiles (Ochs et al., 2018).

The empirical profile is consistent with the theoretical motivation. In 1-D sine approximation with 50 samples and $x\in[t^\ell,t^{\ell+1}]$7 knots, the Lift-Net converges almost in 100 epochs to a near-optimal piecewise-linear fit, while the ReLU network gets stuck in poor local minima even after 4000 epochs. For the 2-D target $x\in[t^\ell,t^{\ell+1}]$8, the lifted model attains lower RMSE in fewer epochs. In CIFAR-10/100, replacing every ReLU by a scaled lifting layer with $x\in[t^\ell,t^{\ell+1}]$9 yields the lowest test error and test-loss among the compared ME-model variants. On MNIST, lifting reaches the lowest final test-error, approximately $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$0. In BSD68 denoising at $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$1, Lift-46 equals or slightly exceeds DnCNN-S and converges more stably over epochs (Ochs et al., 2018).

2. Geometric lifting in human pose estimation

In monocular 3D human pose estimation, “lifting” denotes the reconstruction of 3D structure from 2D keypoints or from keypoints augmented with image cues. LInKs formulates this as a two-step “lift-then-fill” pipeline. With visible keypoints indexed by $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$2, the lifting network produces

$$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$3

and the occlusion-filling network predicts

$$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$4

Its geometric parameterization uses predicted depths $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$5 and

$$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$6

The method partitions the skeleton into legs, torso, left-side, and right-side, and trains an independent lifter on each group, arguing that this reduces long-range keypoint correlations. It also replaces PCA with custom latent-space sampling,

$$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$7

to stabilize normalizing-flow training (Hardy et al., 2023).

The quantitative results show both non-occluded and occluded improvements. On Human3.6M without occlusion and with ground-truth 2D input, LInKs reports PA-MPJPE $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$8 and N-MPJPE $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$9, compared with Wandt et al. at $$\psi(z)=\sum_{i=1}^L z_i t^i, \qquad \psi(\phi(x))=x.$$0 and $$\psi(z)=\sum_{i=1}^L z_i t^i, \qquad \psi(\phi(x))=x.$$1. On MPI-INF-3DHP without occlusion, it reports PA-MPJPE $$\psi(z)=\sum_{i=1}^L z_i t^i, \qquad \psi(\phi(x))=x.$$2, N-PCK $$\psi(z)=\sum_{i=1}^L z_i t^i, \qquad \psi(\phi(x))=x.$$3, and AUC $$\psi(z)=\sum_{i=1}^L z_i t^i, \qquad \psi(\phi(x))=x.$$4. On Human3.6M occlusion benchmarks, filling in 3D improves PA-MPJPE relative to 2D completion for all listed cases: for example, Left arm $$\psi(z)=\sum_{i=1}^L z_i t^i, \qquad \psi(\phi(x))=x.$$5, Right arm $$\psi(z)=\sum_{i=1}^L z_i t^i, \qquad \psi(\phi(x))=x.$$6, Both legs $$\psi(z)=\sum_{i=1}^L z_i t^i, \qquad \psi(\phi(x))=x.$$7, and Full torso $$\psi(z)=\sum_{i=1}^L z_i t^i, \qquad \psi(\phi(x))=x.$$8 (Hardy et al., 2023).

A second line of work argues that lifting from 2D pose alone is limited by depth ambiguity and that image features should be queried explicitly. The “Lifting by Image” framework uses a two-stage pipeline. Stage 1 takes image tokens $$\psi(z)=\sum_{i=1}^L z_i t^i, \qquad \psi(\phi(x))=x.$$9 and pose tokens $\Omega\subset\mathbb R^d$0, couples them through Transformer layers, and predicts a coarse 3D pose. Its Pose-guided Transformer Layer uses

$\Omega\subset\mathbb R^d$1

After Stage 1, the Adaptive Feature Selection Module retains only the top $\Omega\subset\mathbb R^d$2 image tokens, with $\Omega\subset\mathbb R^d$3 by default, and Stage 2 refines the 3D pose using only those retained tokens (Zhou et al., 2023).

The model is explicitly motivated by generalization. The Pose-guided layer reduces background attention from approximately $\Omega\subset\mathbb R^d$4 to approximately $\Omega\subset\mathbb R^d$5 on Human3.6M and from approximately $\Omega\subset\mathbb R^d$6 to approximately $\Omega\subset\mathbb R^d$7 on MPI-INF-3DHP. On Human3.6M under Protocol 1 with CPN 2D detections, it reports average MPJPE $\Omega\subset\mathbb R^d$8 mm; with ground-truth 2D input, average MPJPE $\Omega\subset\mathbb R^d$9 mm; under Protocol 2 with CPN detections, average P-MPJPE $\phi(x)$0 mm. On MPI-INF-3DHP, trained on Human3.6M only, it reports PCK $\phi(x)$1, AUC $\phi(x)$2, and MPJPE $\phi(x)$3 mm (Zhou et al., 2023).

3. Lifting as cut strengthening and convexification in optimization

In integer programming, lifting is a method for strengthening valid inequalities by assigning coefficients to variables not present in a base cut. For the binary knapsack polytope

$\phi(x)$4

a minimal cover $\phi(x)$5 gives the base inequality

$\phi(x)$6

and a lifted cover cut has the form

$\phi(x)$7

The classical GNS sequence-independent lifting constructs a superadditive $\phi(x)$8 from the Balas lifting function. A correction to an overbroad prior generalization shows that the admissible symmetric interpolant must be linear,

$\phi(x)$9

The resulting piecewise-constant lifting is the $d+1$0 case: $d+1$1 Because $d+1$2 is piecewise constant, the coefficient of a variable outside the cover is obtained by interval lookup only, and the paper states that PC lifting runs in $d+1$3 time once the $d+1$4 and $d+1$5 arrays are built (Prasad et al., 2024).

The theoretical contribution is not limited to computability. Broad sufficient conditions are derived under which PC lifting yields a facet-defining cut, including the requirement that at least three variables outside the cover lie in $d+1$6. The computational study, implemented in C++ as a callback within CPLEX v.20.1, reports that on weakly correlated knapsacks with contiguous covers, PC solved $d+1$7 instances, compared with GNS at $d+1$8 and default CPLEX at $d+1$9; on uncorrelated knapsacks, PC solved $f_\theta(x)=\langle \theta,\phi(x)\rangle$0 versus GNS at $f_\theta(x)=\langle \theta,\phi(x)\rangle$1; on Chvátal instances with heaviest-cover cuts, PC and smart lifting solved all $f_\theta(x)=\langle \theta,\phi(x)\rangle$2 instances with tree sizes up to an order of magnitude smaller than GNS or CPLEX’s own cover cuts; and tuned cut limits regained a $f_\theta(x)=\langle \theta,\phi(x)\rangle$3–$f_\theta(x)=\langle \theta,\phi(x)\rangle$4 speedup over default CPLEX in several families (Prasad et al., 2024).

For separable bipartite bilinear programs, lifting is used to derive valid convex inequalities for

$f_\theta(x)=\langle \theta,\phi(x)\rangle$5

Sequential lifting of a seed inequality requires solving a non-convex nonlinear optimization problem for each variable pair. To avoid this, a framework based on subadditive approximations of lifting functions constructs a sequence-independent lift. In the minimal-cover case, the bilinear cover inequality is second-order cone representable, and the relaxation $f_\theta(x)=\langle \theta,\phi(x)\rangle$6 it defines satisfies

$f_\theta(x)=\langle \theta,\phi(x)\rangle$7

so optimizing any nonnegative linear objective over $f_\theta(x)=\langle \theta,\phi(x)\rangle$8 is within factor $f_\theta(x)=\langle \theta,\phi(x)\rangle$9 of the true optimum over $f_\theta(x)=\theta\,\phi(x)$0. The same work develops a two-slope subadditive upper bound $f_\theta(x)=\theta\,\phi(x)$1 and a closed-form lifted bilinear cover inequality for general separable bipartite bilinear sets with box constraints (Gu et al., 2021).

Exact sequential lifting for the binary knapsack set has also been revisited algorithmically. A recent method replaces the classical $f_\theta(x)=\theta\,\phi(x)$2 dynamic-programming array by a dominance list of non-dominated states

$f_\theta(x)=\theta\,\phi(x)$3

with $f_\theta(x)=\theta\,\phi(x)$4 and $f_\theta(x)=\theta\,\phi(x)$5, so that $f_\theta(x)=\theta\,\phi(x)$6 is recovered from breakpoints rather than a full capacity grid. The method preserves scale invariance, handles non-integer coefficients, and admits a reduction rule retaining only

$f_\theta(x)=\theta\,\phi(x)$7

In the reported experiments, for $f_\theta(x)=\theta\,\phi(x)$8 and $f_\theta(x)=\theta\,\phi(x)$9 increasing from $\theta$0 to $\theta$1, array-DP time grew from $\theta$2 s to $\theta$3 s, whereas the list method grew from $\theta$4 s to $\theta$5 s; with the reduction method enabled, those same settings gave roughly $\theta$6 s list time; and embedding the exact list-DP into HiGHS solved $\theta$7 more instances out of $\theta$8 and cut average solve time from $\theta$9 s to $x\in[t^\ell,t^{\ell+1}]$00 s on the affected knapsack cuts (Wang et al., 26 Feb 2026).

4. Function-space and operator lifting in systems and control

In nonlinear systems identification, Koopman-based lifting replaces direct estimation of a nonlinear vector field by linear estimation in a lifted space of observables. Given dictionary functions $x\in[t^\ell,t^{\ell+1}]$01, one defines

$x\in[t^\ell,t^{\ell+1}]$02

forms snapshot matrices from $x\in[t^\ell,t^{\ell+1}]$03 and $x\in[t^\ell,t^{\ell+1}]$04, and estimates a finite-dimensional Koopman operator either by

$x\in[t^\ell,t^{\ell+1}]$05

or, equivalently, by least squares. The infinitesimal generator is then approximated as

$x\in[t^\ell,t^{\ell+1}]$06

and the coefficients of the vector field are recovered by linear regression in the chosen library. A dual method instead works in a $x\in[t^\ell,t^{\ell+1}]$07 sample space, estimates $x\in[t^\ell,t^{\ell+1}]$08, and then reconstructs $x\in[t^\ell,t^{\ell+1}]$09 at data points. The method is described as robust to noise, well-suited to low sampling rate datasets, and applicable to unstable, chaotic, and open systems; the Lorenz example with $x\in[t^\ell,t^{\ell+1}]$10, $x\in[t^\ell,t^{\ell+1}]$11, $x\in[t^\ell,t^{\ell+1}]$12, and $x\in[t^\ell,t^{\ell+1}]$13 reports normalized RMSE approximately $x\in[t^\ell,t^{\ell+1}]$14 over $x\in[t^\ell,t^{\ell+1}]$15 realizations (Mauroy et al., 2016, Mauroy et al., 2017).

In sampled-data control, lifting replaces pointwise state updates by trajectory-valued states over one sampling interval. For the nonlinear continuous-time plant $x\in[t^\ell,t^{\ell+1}]$16, $x\in[t^\ell,t^{\ell+1}]$17 and sampling period $x\in[t^\ell,t^{\ell+1}]$18, the lifted input, output, and state are

$x\in[t^\ell,t^{\ell+1}]$19

The resulting discrete-time system has function-space state and input spaces, and the key claim is that the linear notion of direct feedthrough does not carry over to the nonlinear case; instead, one must further lift the entire state trajectory (Yamamoto et al., 16 Sep 2025). A model-predictive control instantiation of this idea uses fast-sample fast-hold approximations, Runge–Kutta integration, and quadrature inside the NMPC problem. In the reported simulations, lifted NMPC reduced Van der Pol settling time from approximately $x\in[t^\ell,t^{\ell+1}]$20 s to approximately $x\in[t^\ell,t^{\ell+1}]$21 s and reduced the $x\in[t^\ell,t^{\ell+1}]$22-tracking error over the first $x\in[t^\ell,t^{\ell+1}]$23 s by approximately $x\in[t^\ell,t^{\ell+1}]$24; for the inverted pendulum on a cart, lifted NMPC reduced settling from approximately $x\in[t^\ell,t^{\ell+1}]$25 s to approximately $x\in[t^\ell,t^{\ell+1}]$26 s and reduced overshoot by $x\in[t^\ell,t^{\ell+1}]$27, while a multi-rate lifted scheme with $x\in[t^\ell,t^{\ell+1}]$28 succeeded at $x\in[t^\ell,t^{\ell+1}]$29 s where both conventional and single-rate lifted NMPC failed (Gerdpratoom et al., 10 Jan 2025).

Projected Evolutionary Lifting extends the same general idea to infinite-dimensional stochastic control. With a selection operator $x\in[t^\ell,t^{\ell+1}]$30 and weighted path space $x\in[t^\ell,t^{\ell+1}]$31, the lifting operator is

$x\in[t^\ell,t^{\ell+1}]$32

where $x\in[t^\ell,t^{\ell+1}]$33. Liftable functions are those factoring through $x\in[t^\ell,t^{\ell+1}]$34, and under Hypothesis 3.1 the associated semigroup enjoys partial smoothing in the $x\in[t^\ell,t^{\ell+1}]$35-direction. The main existence-and-uniqueness theorem then states that, when the Hamiltonian $x\in[t^\ell,t^{\ell+1}]$36 is concave, for every discount $x\in[t^\ell,t^{\ell+1}]$37 and $x\in[t^\ell,t^{\ell+1}]$38 there exists a unique mild solution

$x\in[t^\ell,t^{\ell+1}]$39

of the lifted stationary HJB, characterized by

$x\in[t^\ell,t^{\ell+1}]$40

thereby removing the earlier restriction to sufficiently large discount factors (Bolli et al., 6 May 2026).

A related optimization-oriented use appears in rigid-body optimal control with contacts. There, lifting reintroduces acceleration, contact forces, and control torques as explicit optimization variables, while inverse dynamics and contact-frame acceleration are imposed as equality constraints. Local elimination of $x\in[t^\ell,t^{\ell+1}]$41 and their dual variables condenses the Newton–KKT system so that the per-iteration cost is almost identical to the conventional non-lifted formulation, yet the convergence speed exceeds twice that of the non-lifted approach. On the ANYmal quadruped, the lifted DMS-LCD method converged in $x\in[t^\ell,t^{\ell+1}]$42–$x\in[t^\ell,t^{\ell+1}]$43 iterations with total solve time approximately $x\in[t^\ell,t^{\ell+1}]$44 s, compared with DMS-CD at $x\in[t^\ell,t^{\ell+1}]$45–$x\in[t^\ell,t^{\ell+1}]$46 iterations and approximately $x\in[t^\ell,t^{\ell+1}]$47 s, DMS-ID at approximately $x\in[t^\ell,t^{\ell+1}]$48 s, FDDP at approximately $x\in[t^\ell,t^{\ell+1}]$49 s, and iLQR at approximately $x\in[t^\ell,t^{\ell+1}]$50 s (Katayama et al., 2021).

5. Algebraic, combinatorial, and semantic lifting

In algebraic statistics, Rauh and Sullivant study lifting of Markov bases and Gröbner bases along a surjective lattice homomorphism $x\in[t^\ell,t^{\ell+1}]$51. Their three-step lifting algorithm first computes a kernel Gröbner basis in $x\in[t^\ell,t^{\ell+1}]$52, then a projected-fiber Gröbner basis for the image family, and finally lifts each projected move $x\in[t^\ell,t^{\ell+1}]$53 back to finitely many moves $x\in[t^\ell,t^{\ell+1}]$54 with $x\in[t^\ell,t^{\ell+1}]$55. The main lifting theorem states that

$x\in[t^\ell,t^{\ell+1}]$56

is a Gröbner basis for the original family. The framework is applied to toric fiber products, where lifted moves from the factors are glued together; under normality assumptions it also yields finiteness results for iterated toric fiber products and uniform degree bounds for families such as the star-graphs $x\in[t^\ell,t^{\ell+1}]$57 and $x\in[t^\ell,t^{\ell+1}]$58 with fixed vertex-sizes in the first part (Rauh et al., 2014).

In coded caching, lifting is a recursive construction on Placement Delivery Arrays. A base PDA is expanded by replacing its entries with constituent PDAs satisfying Blackburn-compatibility or partial Blackburn-compatibility. The odd-tiling construction produces Blackburn-compatible $x\in[t^\ell,t^{\ell+1}]$59 PDAs for odd $x\in[t^\ell,t^{\ell+1}]$60, and lifting the $x\in[t^\ell,t^{\ell+1}]$61-regular PDA $x\in[t^\ell,t^{\ell+1}]$62 by this family yields a $x\in[t^\ell,t^{\ell+1}]$63-regular PDA with parameters

$x\in[t^\ell,t^{\ell+1}]$64

The same paper further shows that the Maddah-Ali–Niesen PDA $x\in[t^\ell,t^{\ell+1}]$65 and Shangguan et al. Construction I can be expressed as lifting constructions under the generalized compatibility notion (Aravind et al., 2023).

Game theory uses “lifting” in a still different sense. One-to-two-player lifting studies when memory guarantees for one-player graph games imply memory guarantees for two-player zero-sum games. If the one-player memory complexity satisfies $x\in[t^\ell,t^{\ell+1}]$66 and

$x\in[t^\ell,t^{\ell+1}]$67

then every payoff $x\in[t^\ell,t^{\ell+1}]$68 satisfies $x\in[t^\ell,t^{\ell+1}]$69. In particular, if $x\in[t^\ell,t^{\ell+1}]$70 on some infinite subsequence, then there is a uniform finite bound on two-player memory. The same work provides the exact barrier by giving a payoff $x\in[t^\ell,t^{\ell+1}]$71 with $x\in[t^\ell,t^{\ell+1}]$72 but $x\in[t^\ell,t^{\ell+1}]$73, so linear one-player memory growth can destroy the lifting phenomenon (Kozachinskiy, 2021).

Scientific computing introduces yet another meaning: stepwise semantics lifting. For LLM-generated FFT kernels, the proposed route is icode $x\in[t^\ell,t^{\ell+1}]$74 $x\in[t^\ell,t^{\ell+1}]$75-SPL $x\in[t^\ell,t^{\ell+1}]$76 SPL $x\in[t^\ell,t^{\ell+1}]$77 high-level FFT specification. Symbolic execution identifies gather/scatter structure such as

$x\in[t^\ell,t^{\ell+1}]$78

which is then matched to known SPL operators such as $x\in[t^\ell,t^{\ell+1}]$79 and recursively assembled into the Cooley–Tukey form. The paper reports preliminary success in lifting GPT-generated FFT code to high-level specifications using an extended SPIRAL framework with symbolic execution and theorem proving (Zhang et al., 15 Jan 2025).

A symmetry-based variant appears in structured convolutional models. There, lifting detects automorphisms of a computation graph, partitions nodes into orbits, and compresses the graph without loss of information. The paper states that detecting symmetries by partition-refinement takes $x\in[t^\ell,t^{\ell+1}]$80, building the compressed graph takes $x\in[t^\ell,t^{\ell+1}]$81, and one forward/backward pass is reduced from $x\in[t^\ell,t^{\ell+1}]$82 on the original graph to $x\in[t^\ell,t^{\ell+1}]$83 on the lifted graph. Reported empirical reductions are $x\in[t^\ell,t^{\ell+1}]$84–$x\in[t^\ell,t^{\ell+1}]$85 on molecule tasks and $x\in[t^\ell,t^{\ell+1}]$86–$x\in[t^\ell,t^{\ell+1}]$87 on knowledge-graph tasks, with speedups up to approximately $x\in[t^\ell,t^{\ell+1}]$88 on small molecule tasks and $x\in[t^\ell,t^{\ell+1}]$89–$x\in[t^\ell,t^{\ell+1}]$90 per-epoch cost relative to unlifted frameworks (Sourek et al., 2020).

6. Embodied lifting: robotic feasibility and human lifting styles

A literal, non-metaphorical use of “lifting technique” appears in humanoid robotics. One method for deciding whether a robot can lift a heavy box with unknown physical properties uses three stages. First, the robot builds an offline table of valid quasi-static lifting trajectories over discretized ranges of box weight $x\in[t^\ell,t^{\ell+1}]$91, center-of-mass offset $x\in[t^\ell,t^{\ell+1}]$92, and gripping distance $x\in[t^\ell,t^{\ell+1}]$93. In the NAO experiments these ranges were discretized as $x\in[t^\ell,t^{\ell+1}]$94–$x\in[t^\ell,t^{\ell+1}]$95 N in $x\in[t^\ell,t^{\ell+1}]$96 N steps, $x\in[t^\ell,t^{\ell+1}]$97 and $x\in[t^\ell,t^{\ell+1}]$98–$x\in[t^\ell,t^{\ell+1}]$99 mm in $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$00 mm steps, yielding $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$01 parameter tuples. Second, online interaction estimates a robust gripping position and the physical parameters using lift height, box rotation, center of pressure, and joint torques. Third, the estimated parameters are conservatively mapped to table indices: the smallest discrete $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$02, the smallest $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$03, and the largest $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$04. If the corresponding table entry contains a valid trajectory, the lift is declared feasible (Han et al., 2020).

The reported demonstration on the NAO humanoid used $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$05 random box configurations with $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$06–$$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$07 kg and $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$08–$$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$09 mm. The robot required at most $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$10 lift-probes to locate a stable $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$11 and estimate parameters; feasibility decisions matched the table in $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$12 of trials; and, when feasible, the stored trajectory lifted the box safely with minimum ZMP margin approximately $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$13 mm inside the support polygon and peak torque approximately $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$14 of $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$15 (Han et al., 2020).

Biomechanics uses “lifting technique” to distinguish posture strategies in manual material handling rather than to denote a representational transform. In a motion-capture-driven musculoskeletal study, $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$16 healthy pain-free individuals repetitively lifted a $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$17 kg box using freestyle, squat, and stoop techniques. Stoop lifting was defined by minimal knee flexion and pronounced lumbar flexion, squat lifting by a spine maintained as straight as possible with greater hip and knee flexion, and freestyle by a self-selected comfortable strategy. Across all lumbar segments, stoop lifting showed significantly lower compressive and total loads, with reductions of $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$18 to $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$19 BW relative to freestyle and squat, but higher shear loads of $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$20 to $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$21 BW in T12/L1 to L4/L5 and lower shear at L5/S1 of $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$22 to $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$23 BW. Peak compressive and total loads during squat lifting occurred approximately $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$24 earlier in the lifting cycle than in stoop lifting. Lumbar lordosis range of motion was $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$25 in stoop, compared with $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$26 in freestyle and $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$27 in squat; lifting time was $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$28 s for freestyle, $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$29 s for squat, and $$\phi(x)=\bigl(1-\lambda_\ell(x)\bigr)e^\ell+\lambda_\ell(x)e^{\ell+1}, \qquad \lambda_\ell(x)=\frac{x-t^\ell}{t^{\ell+1}-t^\ell},$$30 s for stoop (Arx et al., 2021).

The biomechanical interpretation is correspondingly nuanced. The study attributes the compressive-load reduction in stoop lifting to a reduced moment arm of the weight relative to L5/S1 when the hips hinge deeply and the torso mass shifts nearer to the load, while the shear increase in upper lumbar segments is linked to greater lumbar flexion. This directly challenges a common simplification that a flexed spine uniformly increases lumbar loading. A plausible implication is that, even in the literal domain of manual lifting, “best” technique is not a scalar notion but a trade-off between compressive load, shear load, execution speed, and individual tolerance (Arx et al., 2021).