Can DTW compare sequences of different lengths?

Yes. Its alignment can associate one element with several elements in the other sequence under the chosen path rules.

Does it reorder events?

The standard monotonic path preserves sequence order; it changes alignment in time rather than arbitrarily rearranging elements.

Is the lowest-cost alignment always meaningful?

No. It is optimal for the specified costs and constraints, which may not adequately represent the application's notion of similarity.

AI Glossary / Natural language and speech / Intermediate

Dynamic time warping

Definition

Dynamic time warping, or DTW, aligns two sequences while allowing differences in their local timing. It finds a low-cost correspondence between sequence elements under specified ordering and movement constraints. In speech processing, the idea can compare patterns spoken at different speeds without requiring every feature frame to match the same time index.

What it means

Two recordings of the same word can have different durations, and individual sounds may be stretched by different amounts. A direct frame-by-frame comparison can therefore report a large mismatch even when the overall pattern is similar.
DTW permits one element in one sequence to align with multiple elements in the other while preserving order. It searches for an alignment path that minimizes a defined total cost.
The result depends on the feature representation, local distance, allowed path steps, and any additional constraints. A low DTW cost means a good match under those choices; it does not automatically establish that two recordings have identical words, meanings, or speakers.

How it works

Build a local cost matrix comparing every allowed pair of sequence elements. In a common basic formulation, a path starts at the first pair, ends at the last pair, and moves forward horizontally, vertically, or diagonally through the matrix.
Dynamic programming computes the minimum accumulated cost to each cell from the permitted predecessor cells. Backtracking through those choices recovers an optimal alignment.
The basic full-matrix algorithm has O(NM) time complexity for sequence lengths N and M when local comparison costs are treated as constant. Practical variants may restrict the path, use different steps, or normalize scores. Those variants must be specified before comparing results.

Three practical examples

A speech-pattern exercise aligns two recordings whose sounds occur in the same order but have different durations.
A music-analysis tool compares feature sequences from performances with different local tempo patterns.
A sensor-analysis prototype compares an observed motion sequence with a template while allowing some timing variation.

Think of it this way

Think of matching two versions of a melody when one performer holds certain notes longer. DTW resembles pairing corresponding moments without forcing identical clock times. The analogy captures temporal flexibility, but the algorithm follows numerical costs and constraints; it does not inherently understand musical notes or spoken language.

Worked example

Consider invented scalar sequences X = [1, 2, 3] and Y = [1, 1, 2, 3]. Use absolute difference as the local cost and the basic forward steps (1,0), (0,1), and (1,1), with both endpoints included.
One valid path pairs positions (1,1), (1,2), (2,3), and (3,4), using one-based indexing. The paired values are (1,1), (1,1), (2,2), and (3,3). Every local cost is zero, so the total is zero.
Because absolute differences cannot be negative, zero is an optimal total cost here. The first value of X is aligned to two successive values of Y, accommodating the repeated initial value without reversing time.
This does not mean the original sequences are identical: their lengths differ. It shows that this DTW formulation can assign zero cost to distinct sequences related by such repetition. Consequently, calling the score a distance does not imply all the properties of a mathematical metric. Different restrictions or penalties could change the result.

Illustrative case study

Fictional case: MotionMatch compares short exercise demonstrations using sensor features. A fixed-index comparison rates a slower demonstration poorly because its motion phases occur later. The team experiments with DTW to separate timing variation from feature mismatch.
Reviewers also inspect the alignment path. An overly flexible path can stretch one short phase across a long region, producing a deceptively low cost. The team therefore documents its path constraints and evaluates examples with both legitimate speed changes and genuinely different movements.
This fictional case illustrates the tradeoff between flexibility and meaningful alignment. It does not claim medical usefulness, measured recognition accuracy, or that DTW can determine whether a movement is safe.

Common misconceptions and mistakes

DTW does not simply shift one sequence by a fixed delay. It can accommodate local timing differences along the sequence.
Do not assume a zero score proves identical raw sequences, or that every DTW variant satisfies metric properties.
Avoid comparing costs from different local distances, normalization rules, or path constraints as if they shared one scale. Also remember that excessive warping can align patterns that the application should regard as different.

Business use

DTW is useful when a project needs an interpretable alignment between ordered patterns and timing variability is expected. It can support template comparisons, exploratory analysis, and diagnostic visualization.
A project review should inspect both the numerical score and representative paths. Teams should specify preprocessing, feature scale, allowed timing distortion, and how scores are compared across lengths. The relevant question is whether permitted warping reflects variation the application intends to tolerate, rather than whether the algorithm can force a low-cost match.

Check your understanding

Two sequences receive zero DTW cost because one repeats a value that the other contains once. Must they have identical samples and lengths?

Answer: No. A permitted many-to-one alignment can make every paired local difference zero despite unequal sequences. The result must be interpreted under the chosen DTW formulation.

Related terms

Dynamic programming
Time series
Template matching
sequence alignment
warping path
local cost
accumulated cost matrix
speech feature

Learn next

Draw the cost matrix for two short sequences and trace a permitted path. Then add a timing constraint and compare how the allowed alignments change. Inspect whether the resulting score and path match the intended application behavior.

Sources

Sources: Meinard Müller and collaborators, FMP notebook, Dynamic Time Warping: https://www.audiolabs-erlangen.de/resources/MIR/FMP/C3/C3S2_DTWbasic.html ; David Deriso and Stephen Boyd, A General Framework for Dynamic Time Warping: https://stanford.edu/~boyd/papers/gdtw.html Deriso and Boyd primary full paper, Background, consulted at https://stanford.edu/~boyd/papers/pdf/gdtw.pdf for monotonicity, endpoints, adjacent steps and excessive warping. The FMP page was unavailable during this review. An independent basic-DTW implementation confirmed zero cost for the original example; this entry does not attribute its discrete recurrence to the paper’s distinct continuous-time formulation.

Previous: Dynamic state Next: Dynamical system