Primal vs Dual Interpolation

• Primal vs dual interpolation distinctly defines schemes where primal methods guarantee step-wise exact data preservation, while dual methods achieve interpolation in the limit with enhanced smoothness.
• They apply different algebraic frameworks, with primal approaches using simple mask constraints and dual approaches relying on Poisson summation and trigonometric identities.
• In matrix completion, primal formulations optimize low-rank factors for scalability, whereas dual methods employ Lagrange multipliers for robust constraint enforcement, with hybrid methods balancing both advantages.

Primal and dual interpolation are fundamental concepts arising in both stationary subdivision schemes and modern large-scale data interpolation, such as seismic data recovery and low-rank matrix completion. These approaches are distinguished by their theoretical foundations, algebraic constraints, algorithmic structures, and the trade-offs encountered in practical applications.

1. Definitions and Problem Formulations

Stationary subdivision schemes of arity $m\geq2$, characterized by finitely supported masks $\{a_k\}$ and a shift parameter $\tau$, generate basic limit functions $\varphi$ through the refinement equation

$$\varphi(x) = \sum_{k\in\Bbb Z} a_k\,\varphi(m\,x - k + \tau), \qquad x \in \Bbb R.$$

Primal interpolation corresponds to schemes with $\tau=0$ and symmetric support $[-k_r, k_r]$, while dual interpolation corresponds to $\tau=1/2$ and support $[1-k_r, k_r]$. In the context of constrained matrix factorization for data interpolation, primal and dual describe whether optimization proceeds over direct ("primal") variables—typically low-rank factors—or over dual multipliers associated with constraints.

The interpolatory property requires the basic limit function to satisfy $\varphi(n)=\delta_{0,n}$ for $n\in\Bbb Z$. Primal interpolatory schemes yield step-wise interpolation: data at integer sites are exactly retained at every subdivision level. Dual interpolatory schemes produce limit functions that interpolate at integer sites only in the infinite limit, not at each subdivision level (Romani et al., 2019).

In matrix completion for data interpolation, primal formulations minimize the regularizer (e.g., nuclear norm via factorization) subject to a residual constraint, while dual (level-set) approaches maximize dual objectives with respect to dual variables enforcing data constraints (Kumar et al., 2016).

2. Algebraic Characterizations

Subdivision Schemes

For univariate stationary subdivision, primal interpolatory masks $\{a_k\}$ must enforce

$a_{mk} = \delta_{k,0}, \quad k\in\Bbb Z$

which, via sub-symbols $A_n(z)$, is equivalent to

$A_0(z) \equiv \frac{1}{m}, \qquad |z|=1.$

This entails that at each subdivision stage, input data at original lattice points are exactly preserved.

Dual interpolatory masks $\{b_k\}$ have even-length support and do not generally satisfy $b_{mk} = \delta_{k,0}$. The interpolation property is imposed through more intricate trigonometric polynomial identities derived from Poisson summation. For odd and even arity $m$, these identities (see equations (3.4) and (3.6) in (Romani et al., 2019)) provide necessary and sufficient conditions for dual interpolation by enforcing prescribed values at nodes of a sub-lattice and are a direct consequence of the Fourier refinement equation and Poisson summation formula.

Matrix Completion

In matrix interpolation, the primal problem typically minimizes a convex regularizer (such as nuclear norm via matrix factorization), while the dual approach works with the Lagrangian and dual variables. The canonical primal, factorized formulation for observed set $\Omega$ is

$$\min_{L, R} \tfrac{1}{2}(\|L\|_F^2 + \|R\|_F^2) \quad \text{s.t.} \quad \|P_\Omega(LR^H) - b\|_F \leq \eta.$$

The dual formulation introduces dual variables $Y$ representing Lagrange multipliers for the data fidelity constraint, and maximizes a concave dual objective over $Y$. Switching between objective and constraint via root-finding on a value function $v(\tau)$ is a standard primal–dual trick (Kumar et al., 2016).

3. Construction Methodologies

For dual subdivision, construction proceeds as follows:

1. Select half-integer samples: Prescribe $\varphi\bigl(\frac{2k+1}{2}\bigr)$ (often from known primal schemes).
2. Set up the linear system: Evaluate dual characterization equations at multiple frequencies to assemble a matrix–vector system.
3. Enforce convergence and symmetry: Add constraints $\sum_k a_{mk+\gamma} = 1$ for all $\gamma$ and symmetry $a_{-k} = a_{k+1}$.
4. (Optionally) Enforce polynomial reproduction: Factor the symbol to guarantee polynomial reproduction of desired degree.
5. Solve for the mask: The solution yields a dual interpolatory scheme with desired properties (Romani et al., 2019).

In primal–dual matrix interpolation schemes, alternating block-coordinate minimization is performed: in each outer iteration, one alternately solves for $L$ and $R$ in constrained convex subproblems, each reduced via primal–dual splitting (e.g., Chambolle–Pock algorithm), leading to memory and computational efficiencies for large data (Kumar et al., 2016).

4. Comparative Properties and Trade-offs

The key differentiators between primal and dual interpolatory strategies are summarized below:

Aspect	Primal Interpolation	Dual Interpolation
Interpolation behavior	Step-wise, exact at each level	Only in the limit, not at intermediate levels
Mask characteristics	Minimal support, easy construction ($A_0\equiv 1/m$)	Even-length support, supports at half-integers, more flexible
Algebraic structure	Direct condition on mask indices	Poisson-summation/trigonometric polynomial identities
Smoothness	Lower for fixed arity/reproduction degree	Can be higher for same arity, support often slightly longer
Shape control	Limited	Increased, parametric families possible
Polynomial reproduction	Often tightest for given support	Trade-off with smoothness possible via parameter tuning

For example, in ternary cubic-reproducing cases, a dual scheme achieves $C^{2.2760}$ regularity on support $[-3.25,3.25]$, whereas the primal (Dubuc–Deslauriers) scheme achieves $C^{1.8173}$ on $[-2,2]$ (Romani et al., 2019). Dual quinary schemes exhibit $C^{2.20}$ with degree-2 reproduction, faster convergence, and slightly longer support, whereas the corresponding primal binary 5-point achieves $C^{1.415}$ with degree 4.

In low-rank matrix interpolation, the primal factorized approach avoids $nm$-dimensional SVDs by operating on $r(n+m)$ variables, making it more scalable for massive datasets. The dual/level-set approach provides global convexity and monotonic control on constraints but at the expense of higher computational overhead—especially when the data volume is large (Kumar et al., 2016).

5. Algorithmic Strategies: Combined Primal–Dual Approaches

Hybrid methods leverage the advantages of both primal and dual perspectives. In seismic data interpolation, an alternating primal–dual approach:

• Uses factorization for memory and computational efficiency
• Employs dual variables and primal–dual splitting to handle residual constraints efficiently
• Maintains convexity within each block subproblem
• Achieves superior computational performance and scalability compared to level-set schemes (Kumar et al., 2016).

Block-coordinate updates alternate between $L$ and $R$, each solved with a matrix-free primal–dual splitting method. This approach yields $O(1/T)$ convergence for each convex subproblem and empirical wall-clock speedup of 2× over level-set approaches, with uniform cost per frequency slice.

A plausible implication is that such alternating methods benefit from the scalability of the primal approach and the robust constraint handling of the dual, making them particularly suitable for high-dimensional or distributed settings.

6. Guidelines for Method Selection

Selection between primal, dual, and joint approaches depends on application needs:

• Primal subdivision/interpolation: Preferable when step-wise exactness and minimal support are critical, and when algebraic structure should remain simple.
• Dual subdivision/interpolation: Advantageous for higher smoothness, flexible trade-off between support and reproduction, and when parametric control is desired.
• Primal, dual, and primal–dual methods in matrix interpolation:
  • Pure primal (factorized) excels for very large, low-rank matrices using matrix-free operators and parallelization.
  • Pure dual/level-set is suited where convex guarantees and strict control on the nuclear-norm are essential, and for moderate-sized problems with efficient SVDs.
  • Combined approaches are optimal for scalable, distributed, or memory-intensive contexts demanding both speed and robust constraint handling (Romani et al., 2019, Kumar et al., 2016).

Guidelines thus emphasize the tension between interpolation exactness during refinement versus attainable regularity and freedom in design. For maximum smoothness or for tuning complex shape/reproduction trade-offs, dual or hybrid schemes are preferred. If operational simplicity and step-wise data preservation are paramount, primal schemes remain standard.