What is positive definiteness?

“Positive definite” sounds as though it ought to mean something simple.

Perhaps:

definitely positive.

Or at least:

never negative.

It means neither of those things.

A positive-definite function can actually take negative values.

And a function that stays positive everywhere can fail to be positive-definite.

So what on Earth is the “positive” referring to?

The answer is not the value of the function at one point.

It is the behaviour of whole collections of values acting together.

That turns out to be an extraordinarily useful idea.

Start with something that cannot be negative

Take two ordinary real numbers, $a$ and $b$.

Whatever they are,

(A square cannot be negative, whatever numbers we put into it.)

Expand the square:

(One term inside the expression can be negative, but the whole combination is still forced to be zero or above.)

The individual pieces do not all have to be positive.

If $a$ and $b$ have opposite signs, the middle term $2ab$ is negative.

But the whole expression cannot be negative because it came from a square.

That is our first useful intuition.

Positive definiteness is often not about demanding that every ingredient be positive.

It is about demanding that a certain combined quantity can never become negative.

From a square to a matrix

The expression

can be represented by the matrix

If matrices are unfamiliar, you do not need to manipulate this one. For now, think of it simply as a compact table containing the numbers that determine how $a$ and $b$ interact.

The question is:

for which values of $r$ is

for every possible choice of $a$ and $b$?

The answer is

($r$ itself is allowed to be negative. What matters is that no choice of $a$ and $b$ can make the whole quadratic expression negative.)

For example, take

Then the matrix contains negative entries:

Yet the corresponding quadratic expression never becomes negative.

So already we have something slightly surprising:

negative ingredients can live inside a positive-definite structure.

POINTWISE POSITIVITY              POSITIVE DEFINITENESS
check one value f(x)              choose many points + weights
ask: f(x) >= 0?                   combine every pair together
value-by-value test               whole-system consistency test

Positive definiteness is about combinations

For a matrix $M$, we can form a quadratic expression

Here $\mathbf c$ means a list of weights, such as $(c_1,c_2,c_3)$.

(Choose any weights you like, let the matrix combine them, and look at the final number.)

If

for every possible $\mathbf c$, the matrix has a non-negative quadratic form.

For matrices, the usual term for the “zero is allowed” version is positive semidefinite. A positive-definite matrix normally uses the stricter condition $>0$ for every non-zero vector.

For functions, however, the traditional phrase positive-definite function commonly uses the non-negative version.

The terminology is not especially beginner-friendly.

The central idea is:

No permitted weighted combination is allowed to produce a negative quadratic quantity.

Now replace the matrix with relationships

Suppose we choose some points

A function $f$ can describe the relationship between any two of them through their separation:

We can arrange all those pairwise relationships in a matrix.

If matrices still feel unfamiliar, that is fine. Imagine a table: every row chooses one point, every column chooses another, and each cell records the relationship between that pair.

Positive definiteness asks whether every finite table of this kind behaves properly when we combine its entries with arbitrary weights.

For real-valued examples, the condition can be written

(Choose any finite collection of points, give them any real weights, combine every pair, and the final total must never be negative.)

The full general definition allows complex weights and uses a complex conjugate. Nothing important in the intuition changes: the entire quadratic combination must still come out non-negative.

A kernel is a relationship machine

A slightly more general object is called a kernel.

A kernel takes two inputs:

You can think of it, loosely, as assigning a relationship between $x$ and $y$.

It might encode similarity.

Or correlation.

Or overlap.

Or some other interaction.

Given points $x_1,\ldots,x_n$, we form the table

A positive-definite kernel is one for which every such finite table has the required non-negative quadratic behaviour.

A function of a difference,

is a particularly important special case.

This language appears in harmonic analysis, probability, statistics, signal processing, physics and machine learning.

The same structural idea keeps resurfacing.

Here is the surprising part

Consider

Cosine certainly does not stay non-negative.

For example,

(Cosine reaches minus one, so it definitely fails ordinary pointwise non-negativity.)

If positive-definite meant “positive everywhere”, cosine would therefore be disqualified immediately.

But cosine is positive-definite.

We can see why without an advanced theorem.

Take any real points $x_1,\ldots,x_n$ and any real weights $c_1,\ldots,c_n$.

Consider

Using the identity

the entire expression becomes

(The intimidating double sum has turned into two ordinary squares added together.)

And two squares added together cannot be negative.

So

can take negative values while still being positive-definite.

That is probably the single most useful example for separating the two ideas.

And the reverse surprise also happens

Now consider

This function is positive everywhere:

Surely that ought to make it positive-definite?

No.

Take two points,

and choose the weights

The positive-definiteness test becomes

Since

we get

(Every individual value of $1+x^2$ is positive, yet this perfectly legitimate combined test gives a negative result.)

Therefore $1+x^2$ is positive everywhere but is not positive-definite.

So pointwise positivity and positive definiteness are not simply weaker and stronger versions of one another.

They are different kinds of positivity.

Why would anyone want such a condition?

Because it appears naturally whenever numbers describe relationships.

Imagine that $K(x,y)$ describes how similar two things are.

Or how strongly two measurements are correlated.

Or how much two signals overlap.

The individual relationships can sometimes be negative.

Negative correlation, for example, is perfectly meaningful.

But when all the relationships come from a genuine underlying system, certain total quadratic quantities cannot be negative.

That is exactly the kind of consistency that positive definiteness captures.

Correlation gives a beautiful example

Suppose $X(t)$ is a signal or random process and $R(\tau)$ measures how strongly it correlates with a shifted version of itself.

A collection of shifted copies gives pairwise correlations such as

Now choose arbitrary weights $c_j$ and combine those shifted copies.

The corresponding quadratic correlation quantity has the form

(We are asking about the total correlation or energy-like quantity of an arbitrary weighted mixture of shifted signals.)

Under the usual correlation construction, that total is ultimately an average square or squared magnitude, so it cannot be negative.

That is why autocorrelation functions naturally have positive-definite structure.

The mathematics is not imposing a decorative extra rule.

The positivity is telling us that all the pairwise relationships are mutually consistent with an underlying signal or random process.

Variance tells the same story

Probability provides another familiar example.

A covariance matrix contains the pairwise covariances between random variables.

Some of those covariances may be negative.

That simply means two quantities tend to move in opposite directions.

But take any weighted combination of the random variables.

Its variance cannot be negative.

So the covariance matrix must have non-negative quadratic form.

Again, individual entries need not all be positive.

The positivity belongs to the whole structure.

A useful informal translation

Pointwise positivity asks:

Is this individual value non-negative?

Positive definiteness asks something more like:

Can all these pairwise relationships coexist without producing an impossible negative square, variance or energy?

That is why positive definiteness feels more global.

It tests collections of points and their relationships rather than inspecting one value at a time.

Then Fourier analysis enters the story

This is where the idea becomes especially beautiful.

Fourier analysis gives two different ways of describing many mathematical objects: one in terms of position, separation or time, and another in terms of frequency.

A theorem called Bochner's theorem connects this directly to positive definiteness.

For a continuous positive-definite function on ordinary Euclidean space, the function can be represented as the Fourier transform of a finite non-negative measure.

For suitably well-behaved integrable functions, the takeaway is especially intuitive:

positive definiteness in one description corresponds to non-negative spectral content in the Fourier description.

(A complicated “all points and all weights” condition can sometimes be checked by looking for negative content on the frequency side.)

Why that should feel familiar

In signal processing, a genuine power spectral density cannot contain negative power.

Meanwhile, the corresponding autocorrelation function has positive-definite structure.

So the two statements

correlation is positive-definite

and

the spectral measure is non-negative

are Fourier partners.

This is one reason positive definiteness appears so naturally whenever correlations, spectra and Fourier transforms meet.

It is also exactly why the idea appears in Riemann Console.

The Riemann Console connection

#

Positive definiteness matters to Riemann Console because an earlier project route translated a structural positivity question into a spectral sign condition. That made the route unusually falsifiable: a reliable negative spectral value would obstruct the stronger positive-definiteness target.

#

Hostile numerical testing eventually found such an obstruction, and the programme retired that sufficient route. The result did not disprove RH; it showed that one stronger structural condition was too demanding.

#

The earlier public version of this explainer reproduced the project's exact internal construction. Following a strategic-disclosure review on 19 September 2026, those definitions, transform identities, coordinates and numerical values are no longer reproduced here. The mathematical lesson does not require them: pointwise sign, positive definiteness and spectral positivity are related concepts, but they are not interchangeable.

#

For the bounded research record, see MIL-RH-0003.

#

Why positive definiteness is powerful

When an object really is positive-definite, a great deal of structure can become available.

It may behave like a correlation.

It may admit a spectral representation by a non-negative measure.

Matrices built from sampled points inherit non-negative quadratic forms.

Probabilistic interpretations may become possible.

Geometric or Hilbert-space representations may appear.

This is why mathematicians pay attention when positive definiteness shows up.

It can indicate that a complicated-looking function is secretly behaving like an inner product, covariance, correlation or spectrum.

There is a geometric picture hiding underneath

Suppose we can associate a vector $\Phi(x)$ with every point $x$ so that

(The kernel value is acting like the dot product, or geometric relationship, between two vectors.)

Then

becomes

(After combining all the pairwise relationships, the answer turns into a squared length.)

And a squared length cannot be negative:

This is perhaps the cleanest conceptual picture of all.

The mysterious quadratic inequality has turned back into geometry.

We have come full circle.

So why is cosine positive-definite?

Now the earlier surprise becomes almost visual.

Associate each real number $x$ with the two-dimensional vector

(Each number $x$ becomes a point on the unit circle.)

Then

(The cosine of the separation between $x$ and $y$ is exactly the inner product of their two unit-circle vectors.)

So the cosine kernel really does arise from ordinary inner products.

That is why it is positive-definite.

Its negative values are not a contradiction.

Two vectors can have a negative inner product.

What cannot be negative is the squared length produced when the whole system is combined.

That is the distinction.

A small glossary

Positive definite

A structural positivity property requiring every permitted finite quadratic combination of an object's pairwise values to be non-negative.

Positive semidefinite matrix

A matrix $M$ for which

for every vector $\mathbf c$.

Quadratic form

An expression in which weights interact pairwise, such as

Kernel

A function $K(x,y)$ that assigns a value to a pair of inputs. Kernels often encode similarity, correlation or interaction.

Positive-definite function

In a translation-invariant setting, a function $f$ for which

is a positive-definite kernel.

Fourier transform

A mathematical transformation that rewrites an object in terms of frequency content.

Bochner's theorem

A theorem connecting continuous positive-definite functions with Fourier transforms of finite non-negative measures.

Spectral measure

A measure describing how an object is distributed across frequency. In Bochner's theorem, its non-negativity is the spectral counterpart of positive definiteness.

The one sentence to remember

Positive definite does not mean that a function is positive everywhere.

It means something more structural:

every finite pattern of its pairwise relationships must combine to give a non-negative quadratic quantity.

That is why a function such as $\cos x$ may dip below zero and still be positive-definite — and why positive definiteness can reveal structure that ordinary positivity cannot see.