For , write to emphasize that this is a quasi-norm, as in the Lq quasi-norm branch. On two coordinates with counting measure, and give , but . At , this is the explicit failure in the right panel of the diagram. The midpoint has , so the corresponding unit sublevel set is not a convex set.
A useful replacement survives: for nonnegative real numbers , . Together with the scalar triangle inequality, this yields
Consequently is a metric on the equivalence classes of measurable functions with finite , even though itself need not be a metric. Do not replace the failed Minkowski inequality by a universal reversed bound for signed or complex functions: would give a zero sum but two positive separate sizes.
Lp space Created 2026-10-14 Updated 2026-11-05
For a measure space and , consists of equivalence classes of measurable functions satisfying
Two representatives are identified when they agree almost everywhere. For , use the essential supremum norm, not an infinite power inside the integral. The quotient is essential: a function supported on a null set has zero integral norm even if it is not identically zero.[ref]
These are Banach spaces; L2 space also has the integral inner product making it a Hilbert space. On with Lebesgue measure, , , belongs to exactly when . This shows why membership depends on the exponent: a function can be integrable while its square is not. For , the same power expression defines a quasi-norm rather than a norm, as explained under Lq quasi-norm.
For vectors or , counting measure turns the Minkowski inequality into
For , replace each displayed norm by the maximum of its coordinate absolute values. Take and . Both vectors have Lp norm one for every , but their sum has L1 norm , Euclidean norm , and L-infinity norm . Thus the same vector addition gives equality at , strict inequality at , and strict inequality at .
By contrast, and give equality for every : absolute homogeneity of a norm yields . At , the three sizes are , , and . Opposite vectors behave very differently: , so the inequality is strict whenever is nonzero. Cancellation is allowed; the bound describes the largest possible total size.
Figure 1.
Strict Minkowski inequality, equality for positive multiples, and quasi-norm failure
. Arrows show , the translated step , and their sum. Left: perpendicular unit vectors give for the Euclidean norm. Middle: and give . Right: and give for the Lq quasi-norm at . Pale curves show the corresponding unit level sets. Details.
Use the stated characterization by the Lq null space property. Raising an Lq quasi-norm comparison to its positive exponent preserves its order, so the hypothesis at says that, for every nonzero and every ,
We prove the corresponding inequality; this is monotonicity of uniform sparse recovery in the exponent.
Fix a nonzero null space vector and arrange its coordinate magnitudes as . Assume . The Lq null space property rules out a nonzero null space vector supported on at most coordinates, so and . Since , the factors are at most for and at least for with . Terms with contribute zero and require no negative power of zero. Thus
The largest coordinates maximize the -power sum on any set of size at most . Its complement therefore has the smallest complementary -power sum. The displayed strict inequality proves the Lq null space property at exponent for every allowed support of a vector. The stated recovery characterization now applies to the Lq quasi-norm at .
If , the measurement constraint already singles out , for every objective; if , only the zero sparse vector needs recovery. These cases do not need a positive threshold. Uniform recovery at implies uniform recovery at every :
It is equivalent to minimize or its th power , because taking the th root is strictly increasing. For , this is an Lq quasi-norm, while gives the L1 norm.
Assume the Lq null space property of order . Let have support of a vector with , and let be any distinct feasible vector. Then . The triangle inequality for complex magnitudes and the subadditivity give
For the same inequality follows directly from the triangle inequality. Since off ,
Thus is the unique minimizer.
Conversely, suppose every -sparse vector is the unique minimizer for its measurements. Fix and any with . The vectors and are distinct and satisfy , since . Uniqueness for this therefore implies , exactly the required strict inequality. Hence
The quantifier is uniform over the entire sparse class. Recovery of one particular signed vector alone would not imply this null space property.