Total variation measures accumulated absolute change rather than just the difference between two endpoint values. For a real or complex sequence , its total variation is
allowing . The triangle inequality gives . Finite variation therefore makes the sequence Cauchy and hence convergent in or .
The converse fails. The sequence tends to zero, but , whose sum diverges. Small late values can still accumulate infinitely much back-and-forth movement.
For a function on an interval, the total variation of a function replaces the sum over consecutive indices by a supremum over all partitions of an interval. For spatial fields, the total variation seminorm on a domain uses the distributional derivative as a measure. Total generalized variation extends that regularization idea to higher-order structure. These settings share cumulative-change control, while their definitions specify different index or spatial domains.
Second-order total generalized variation combines a first derivative with an auxiliary vector field and its symmetric derivative. A standard continuous form is
Here denotes the symmetric distributional derivative and the norms are total variations of the corresponding measures. Discrete variants replace these operators by linear maps and use sums of row norms. Unlike first-order total variation, this regularizer accommodates piecewise-affine behavior. A TGV divergence splitting makes its constrained divergence dual amenable to explicit proximal operators.
For a quadratic data term and divergence maps , the dual constraint can be split through . For a product of Euclidean row balls of radius , its support function is . The resulting saddle coupling is . The Chambolle–Pock algorithm then uses a row-ball projection, a quadratic proximal operator and radial soft thresholding, with no projection onto an intersection involving a divergence operator.

Articles by others on the same topic (1)