Curvature operators and scalar curvature invariants

Curvature operators and scalar curvature invariants


We continue the study of the question of when a pseudo-Riemannain manifold can be locally characterised by its scalar polynomial curvature invariants (constructed from the Riemann tensor and its covariant derivatives). We make further use of alignment theory and the bivector form of the Weyl operator in higher dimensions, and introduce the important notions of diagonalisability and (complex) analytic metric extension. We show that if there exists an analytic metric extension of an arbitrary dimensional space of any signature to a Riemannian space (of Euclidean signature), then that space is characterised by its scalar curvature invariants. In particular, we discuss the Lorentzian case and the neutral signature case in four dimensions in more detail.

1 Introduction

Recently, the question of when a pseudo-Riemannain manifold can be locally characterized by its scalar polynomial curvature invariants constructed from the Riemann tensor and its covariant derivatives has been addressed.

In [1] it was shown that in four dimensions (4d) a Lorentzian spacetime metric is either -non-degenerate, and hence locally characterized by its scalar polynomial curvature invariants, or is a degenerate Kundt spacetime [2]. Therefore, the degenerate Kundt spacetimes are the only spacetimes in 4d that are not -non-degenerate, and their metrics are the only metrics not uniquely determined by their curvature invariants. In the proof of the -non-degenerate theorem in [1] it was necessary, for example, to determine for which Segre types for the Ricci tensor the spacetime is -non-degenerate. By analogy, in higher dimensions it is useful to utilize the bivector formalism for the Weyl tensor. Indeed, by defining the Weyl bivector operator in higher dimensions [3] and making use of the alignment theory [4], it is possible to algebraically classify any tensor (including the Weyl tensor utilizing the eigenbivector problem) in a Lorentzian spacetime of arbitrary dimensions, which has proven very useful in contemporary theoretical physics.

The link between a metric (or rather, when two metrics can be distinguished up to diffeomorphisms) and the curvature tensors is provided through Cartan’s equivalence principle. An exact statement of this is given in [5]; however, the idea is that knowing the Riemann curvature tensor, and its covariant derivatives, with respect to a fixed frame determines the metric up to isometry. It should be pointed out that it may be difficult to actually construct the metric, however, the metric is determined in principle. More precisely, given a point, , a frame , and the corresponding Riemann tensor and its covariant derivatives , then if there exists a map such that:

then there exists an isometry such that (which induces the abovementioned map). Therefore, if we know the components of the Riemann tensor and its covariant derivatives with respect to a fixed frame, then the geometry and, in principle, the metric is determined. The equivalence principle consequently manifests the connection between the curvature tensors and the spacetime metric. The question of whether we can, at least in principle, reconstruct the metric from the invariants thus hinges on the question whether we can reconstruct the curvature tensors from its scalar polynomial curvature invariants, indicated with a question mark in the following figure:

It is to address this question, the curvature operators are particularly useful.

In this paper we continue the study of pseudo-Riemannain manifolds and their local scalar polynomial curvature invariants. In principle, we are interested in the case of arbitrary dimensions and all signatures, although we are primarily interested from a physical point of view in the Lorentzian case and most illustrations will be done in 4d. We also briefly discuss the neutral signature case.

First, we shall discuss the concept of a differentiable manifold being characterised by its scalar invariants in more detail. Suppose we consider the set of invariants as a function of the metric and its derivatives: (the -map); we are then interested in under what circumstances, given a set of invariants , this function has an inverse . For a metric which is -non-degenerate the invariants characterize the spacetime uniquely, at least locally, in the space of (Lorentzian) metrics, which means that we can therefore distinguish such metrics using their curvature invariants and, hence, is unique (up to diffeomorphisms).

We then further discuss curvature operators (and introduce an operator calculus) and their relationship to spacetime invariants. Any even-ranked tensor can be considered as an operator, and there is a natural matrix representation of the operator . In particular, for a curvature operator (such as the zeroth order curvature operator, ) we can consider an eigenvector with eigenvalue ; i.e., . This allows us to introduce the important notion of diagonalisability. In addition, we then introduce the notion of analytic metric extension, which is a complex analytic continuation of a real metric under more general coordinate transformations than the real diffeomorphisms which results in a real bilinear form but which does not necessarily preserve the metric signature. We shall show that if a space, , of any signature can be analytically continued (in this sense) to a Riemannian space (of Euclidean signature), then that spacetime is characterised by its invariants. Moreover, if a spacetime is not characterised by its invariants, then there exists no such analytical continuation of it to a Riemannian space.

Finally, we present a summary of the Lorentzian signature case. All of the results presented here are illustrated in the 4d case. We also discuss the 4d pseudo-Riemannian case of neutral signature (NS space).

1.1 Metrics characterised by their invariants

Consider the continuous metric deformations defined as follows [1].

Definition 1.1.

For a spacetime , a (one-parameter) metric deformation, , , is a family of smooth metrics on such that

  1. is continuous in ,

  2. ; and

  3. for , is not diffeomorphic to .

For any given spacetime we define the set of all scalar polynomial curvature invariants

Therefore, we can consider the set of invariants as a function of the metric and its derivatives. However, we are interested in to what extent, or under what circumstances, this function has an inverse. More precisely, define the function (the -map) to be the function that calculates the set of invariants from a metric . Given a set of invariants, , what is the nature of the (inverse) set ? In order to address this question, we first need to introduce some terminology [1].

Definition 1.2.

Consider a spacetime with a set of invariants . Then, if there does not exist a metric deformation of having the same set of invariants as , then we will call the set of invariants non-degenerate. Furthermore, the spacetime metric , will be called -non-degenerate.

This implies that for a metric which is -non-degenerate the invariants characterize the spacetime uniquely, at least locally, in the space of (Lorentzian) metrics. This means that these metrics are characterized by their curvature invariants and therefore we can distinguish such metrics using their invariants. Since scalar curvature invariants are manifestly diffeomorphism-invariant we can thereby avoid the difficult issue whether a diffeomorphism exists connecting two spacetimes.

In the above sense of -non-degeneracy, a spacetime is completely characterized by its invariants and there is only one (non-isomorphic) spacetime with this set of invariants (at least locally in the space of metrics, in the sense above). However, we may also want to use the notion of characterization by scalar invariants in a different (weaker) sense (see below). In order to emphasise the difference between such cases, we will say that -non-degenerate metrics are characterised by their invariants in the strong sense, or in short, strongly characterised by their invariants (this is the general sense, and if the clarifier ‘strong sense’ is omitted, this meaning is implied).

However, in the definition 3.3 (below) it is useful to regard a pseudo-Riemannian manifold as being characterised by its invariants if the curvature tensors (i.e., the Riemann tensor and all of its covariant derivatives) can be reconstructed from a knowledge all of the scalar invariants. It is clear that -non-degenerate metrics fall into this class, but another important class of metrics will also be included. For example, (anti-)de Sitter space is maximally symmetric and consequently has only one independent curvature component, and also only one independent curvature invariant (namely the Ricci scalar). Therefore, we can determine the curvature tensors of (anti-)de Sitter space by knowing its invariants. In our definition, (anti-)de Sitter space is not -non-degenerate but it will be regarded as being characterised by its invariants.

Therefore, we shall say that metrics that are characterised by their invariants but are not -non-degenerate, are weakly characterised by their invariants. A common factor for metrics being characterised by their invariants is that the curvature tensors, and hence the spacetime, will inherit certain properties of the invariants. In particular, as we will see later, all CSI spacetimes being characterised by their invariants will be automatically be locally homogeneous; however, not all of them are -non-degenerate, like (anti-)de Sitter space.

The definition we will adopt is closely related to the familiar problem of diagonalising matrices. We will use curvature operator to relate our problem to linear algebra, and essentially, we will say that a spacetime is characterised by its invariants if we can diagonalise its curvature operators. In order to make this definition a bit more rigourous, we need to introduce some formal definitons.

2 Definition

Consider an even-ranked tensor . By raising or lowering the indices appropriately, we can construct the tensor with components . This tensor can be considered as an operator (or an endomorphism) mapping contravariant tensors onto contravariant tensors

where is the vectorspace , or, if the tensor possess index symmetries, . Here, since is a vectorspace, there exists a set of basis vectors so that . Expressing in this basis, we can write the operator in component form: . This means that we have a natural matrix representation of the operator .

Note that we can similarly define a dual operator where , mapping covariant tensors onto covariant tensors. In component form the dual operator is which in matrix form can be seen as the transpose of ; i.e., . Consequently, there is a natural isomorphism between the operator and its dual. We can also consider operators mapping mixed tensors onto mixed tensors.

For an odd-ranked tensor, , we can also construct an operator: however, this time we need to consider the tensor product of with itself, , which is of even rank. We can also construct an operator with two odd-ranked tensors and : . In this way is even-ranked and we can lower/raise components appropriately.

Now, consider an operator. If are the components in a certain basis, then this as a natural matrix representation. This is an advantage because now we can use standard results from linear algebra to estabish some related operators.

The utility of these operators when it comes to invariants is obvious since any curvature invariant is always an invariant constructed from some operator. For example, the Kretchmann invariant, , is trivially (proportional to) the trace of the operator . Similarly, an arbitrary invariant, , is the trace of the operator . Therefore, it is clear that we can study the polynomial curvature invariants by studing the invariants of curvature operators.

The archetypical example of a curvature operator is the Ricci operator, , which maps vectors onto vectors; i.e.,

All of the Ricci invariants can be constructed from the invariants of this operator, for example, is the Ricci scalar and .

Another commonly used operator is the Weyl operator, , which maps bivectors onto bivectors:

The Weyl invariants can also be constructed from this operator by considering traces of powers of : .

3 Eigenvalues and projectors

For a curvature operator, , consider an eigenvector with eigenvalue ; i.e., . Note that the symmetry group of the spacetime (of signature ), is naturally imbedded through the tensor products . By considering the eigenvalues of as solutions of the characteristic equation:

which are -invariants. Since the orthogonal group , using the tensor product, acts via a representation , Hence, the eigenvalue of a curvature operator is an -invariant curvature scalar. Therefore, curvature operators naturally provide us with a set of curvature invariants (not necessarily polynomial invariants but derivable from them) corresponding to the set of distinct eigenvalues: . Furthermore, the set of eigenvalues are uniquely determined by the polynomial invariants of via its characteristic equation. The characteristic equation, when solved, gives us the set of eigenvalues, and hence these are consequently determined by the invariants.

We can now define a number of associated curvature operators. For example, for an eigenvector so that , we can construct the annihilator operator:

Considering the Jordan block form of , the eigenvalue corresponds to a set of Jordan blocks. These blocks are of the form:

There might be several such blocks corresponding to an eigenvalue ; however, they are all such that is nilpotent and hence there exists an such that annihilates the whole vector space associated with the eigenvalue .

This implies that we can define a set of operators with eigenvalues or by considering the products

where (as long as for all ). Furthermore, we can now define

where is a curvature projector. The set of all such curvature projectors obeys:


We can use these curvature projectors to decompose the operator :


The operator thus contains all the information not encapsulated in the eigenvalues . From the Jordan form we can see that is nilpotent; i.e., there exists an such that . In particular, if , then is a negative/positive boost weight operator which can be used to lower/raise the boost weight of a tensor [4, 3].

We also note that

Consequently, we have the orthogonal decomposition:


Note that we can achieve this decomposition by just using operators and their invariants.

In linear algebra we are accustomed to the concept of diagonalisable matrices. Similarly, we will call an operator diagonalisable if


In differential geometry we are particularly interested in quantities like the Riemann curvature tensor, . Now, the Riemann tensor naturally defines two curvature operators: namely, the Ricci operator , and the Weyl operator . Let us, at a point , define the tensor algebra (or tensor concomitants) of 0th order curvature operators, ,1 as follows:

Definition 3.1.

The set is a -linear space of operators (endomorphisms) which satisfies the following properties:

  1. Identity: the metric tensor, as an operator, is in .

  2. the Riemann tensor, as an operator, is in .

  3. if , then any tensor contraction and any eigenvalue of is also in .

  4. “Square roots”: if and (as tensors) where is of even rank, then

  5. any operator that can be considered as a tensor product or expressible as functions of elements in is also in .

An element of , will be referred to as a curvature operator of order 0.

Similarly, allowing for contractions of the covariant derivative of the Riemann tensor, , we can define the tensor algebra of 1st order curvature operators, , etc. If we allow for arbitrary number of derivatives, we will simply call it . Hence2:

For example, the following tensor:

is a curvature tensor of order 2 and is consequently an element of .

Of particular interest is when all curvature operators can be determined using their invariants. This means that there is preferred set of operators (namely the projectors) which acts as a basis for the curvature operators. This is related to diagonalisability of operators, and hence:

Definition 3.2.

is called diagonalisable if there exists a set of projectors forming a tensor basis for ; i.e., for every ,

Note that the components may be complex. If an operator is diagonalisable, we can, by solving the characteristic equation, determine the expansion eq. (4). This enables us to reconstruct the operator itself. In this sense, the operator would be characterised by its invariants. Hence, we have the following definition:

Definition 3.3.

A space is said to be characterised by its invariants iff, for every , the set of curvature operators, , is diagonalisable (over ).

As pointed out earlier we can further subdivide this category into spaces that are characterised by their invariants in a strong or weak sense. If, in addition, the space is -non-degenerate, then the space is characterised by its invariants in the strong sense; otherwise its said to be characterised by its invariants in the weak sense.

A word of caution: note that even though we know all the operators of a spacetime that is characterised by its invariants, we do not necessarily know the frame. In particular, by raising an index of the metric tensor, , we get, irrespective of the signature, . This means that for a Lorentzian spacetime we would lose the information of which direction is time. However, there might still be ways of determining which is time by inspection of the invariants (but this is no guarantee). If, say, an invariant can be written as , and , then clearly, is timelike and, consequently, can be used as time. On the other hand, an example of when this cannot be done is the following case, in dimensions, when the only non-zero invariants are the 0th order Ricci invariants:

Then, if this spacetime is characterised by its invariants, we get:

However, there is no information in the invariants if the direction associated with the eigenvalue is a space-like or time-like direction. Consequently, in the definition above we have to keep in mind that these spacetimes are characterised by their invariants, up to a possible ambiguity in which direction is associated with time, and which is space3.

In the following we will reserve the word “spacetime” to the Lorentzian-signature case, while “Riemannian” (space) will correspond to the Euclidean signature case. For the Riemannian case, the operators provide us with an extremely simple proof of the following theorem:

Theorem 3.4.

A Riemannian space is always characterised by its invariants.


In an orthonormal frame, the Riemannian metric is . Therefore, the components of an operator will be the same as those of , possibly up to an overall constant. Therefore, let us decompose into the symmetric, , and antisymmetric part , of :

Since the signature is Euclidean, the operators and will also be symmetric and antisymmetric (as matrices), respectively. A standard result is that we can consequently always diagonalise :


i.e., the operator is diagonalisable. For :


Consequently, is diagonalisable also (with purely imaginary eigenvalues).

So for , then also and , and therefore any curvature operator is characterised by its invariants. This proves the theorem. ∎

We note that this theorem also follows from a group-theoretical perspective [6, 7]; however, the operators provide an alternative (and fairly straight-forward) proof.

Of course, it is known that a Lorentzian spacetime is not necessarily characterised by its invariants in any dimension.4 However, the spacetimes being characterised by their curvature invariants play a special role for certain classes of metrics. For example, we have that:

Proposition 3.5 (VSI spacetimes).

Among the spacetimes of vanishing scalar curvature invariants (VSI), the only spacetimes being characterised by their curvature invariants (weakly or strongly) are locally isometric to flat space.

Note that flat space is a characterised by its invariants only in the weak sense.

Proposition 3.6 (CSI spacetimes).

If a spacetime has constant scalar curvature invariants (CSI) and is characterised by its invariants (weakly or strongly), then it is a locally homogeneous space.

These locally homogeneous spacetimes can be characterised by their invariants in either a weak or strong sense.

Regarding the question of which spacetimes are characterised by its invariants, in the sense defined above, we have that the following conjecture is true:

Conjecture 3.7.

If a spacetime is characterised by its invariants (weakly or strongly), then it is either -non-degenerate, or of type D to all orders; i.e., type D.

There are several results that support this conjecture. First, it seems reasonable that all -non-degenerate metrics are (strongly) characterised by their invariants. In 4 dimensions, this seems clear from the results of [1]. Furthermore, that type D spacetimes are also characterised by its invariants (in a weak sense necessarily) follows from the results of [8]. However, it remains to prove that these are all.

We should also point out a sublety in the definition of what we mean by characterised by the invariants. Now a class of metrics that, in general, does not seem to be characterised by their curvature invariants is the subclass of Kundt metrics [1, 2]:

For sufficiently general functions , , and , these metrics will not be characterised by their invariants as defined above because they will be of type II to all orders (in three dimensions, see the example below). However, the invariants still enable us to reconstruct the metric. In some sense, the metric above is the simplest metric with this set of invariants (however, it is not characterised by its invariants as defined above!).

Example: 3D degenerate Kundt metrics.

Let use study the details of the above example in three dimensions, but generalising to the full class of degenerate Kundt metrics, which can be written:

where and . In 3D the Weyl tensor is zero and so we only need to consider the Ricci tensor which, using the coordinate basis , can be written in operator form:


The eigenvalues are ; consequently, if the Segre type is or . Unfortunately, the coordinate basis is not the canonical Segre basis so the form is not manifest. However, the projection operators are frame independent operators so let us find the various projectors in this case. Since has multiplicity 2, (while has multiplicity 1) the operator must be proportional to the projection operator . By ordinary matrix multiplication we get:


The proportionality constant is . Consequently:

We note that the projection operator is not diagonal in the coordinate basis: however, we can easily verify that .

There are only two projection operators in this case, so can be calculated using . Alternatively, we can define so that the projection operator is given by . Using either way to calculate , we obtain:

To check which Segre type the Ricci tensor has we calculate the nilpotent operator given by the expansion eq. (2):


Therefore, the metric is Segre type in general, while if then it is Segre type .

For the above metric we have


The invariants will only depend on and , and the invariants will determine these functions up to diffeomorphisms. However, in general, this metric will be of Segre type ; therefore, even if , this metric will not be determined by the curvature invariants, as defined above.

4 The operator calculus

Since all of the results so far are entirely point-wise, it is also an advantage to consider the operator calculus; i.e. derivatives of operators. Most useful for our purposes is the Lie derivative. The Lie derivative preserves the order and type of tensors and is thus particularly useful.

Consider a vector field defined on a neighbourhood . We must assume that the operator decomposition, (3), does not change over . This assumption is essential in what follows.

Consider the one-parameter group of diffeomorphisms, , generated by the vector field . Assume that two points and (both in ) are connected via ; i.e., . Since the eigenvalues are scalar functions over , then . Furthermore, by assumption, the eigenvalue structure of does not change over ; consequently, for a eigenvector with eigenvalue we have:

This implies that eigenvectors are mapped onto eigenvectors. Therefore, if the eigenvalue is mapped onto (these are scalar functions determined by the characteristic equation), then the eigenvector is mapped onto the corresponding eigenvector . Then if spans the eigenvectors of eigenvalue , and spans the eigenvectors with eigenvalue , then there must exist, since also preserves the norm, an invertible matrix so that:

If and are the corresponding one-forms, then . For a projector, the eigenvalues are ; therefore, we can write

where the eigenvectors have eigenvalue 1. Therefore:


Therefore, we have the remarkable property that the projectors are Lie transported with respect to any vector field .

For the Lie derivative we thus have the following result:

Theorem 4.1.

Assume that the operator, , has the decomposition (3). Then, the Lie derivative with respect to is


We have


Since , the theorem now follows. ∎

Now clearly, this has some interesting consequences. If, for example, the operator is diagonalisable, i.e., , then is Lie transported if and only if the eigenvalues are Lie transported. This can formulated as follows:

Corollary 4.2.

If an operator, T, is diagonalisable, then, for a vector field, ,

Corollary 4.3.

If a spacetime is characterised by its invariants (weakly or strongly), then if there exists a vector field, , such that

then there exists a set, , of Killing vector fields such that at any point the vector field coinsides with a Killing vector field .

The last corollary enables us to reduce the question of Killing vectors down to the existence of vectors annihilating all the curvature invariants . This may be easier in some cases if the form of the metric is totally unmanageable.


CSI spacetimes:

Consider now a spacetime which has all constant scalar curvature invariants [9, 10, 11]. Assume also that the spacetime is characterised by its invariants, weakly or strongly. Let us now see how we can give an alternative proof that this must be locally homogeneous.

We choose, at any point , a local coordinate system . Using the coordinate frame, we can define we get:

Therefore, for each (which are linearly independent at ), there would be a Killing vector field coinsiding with at . Therefore, the Killing vectors are also linearly independent and span the tangent space, consequently, the spacetime is locally homogeneous.

Kinnersley class I vacuum metrics:

Let us consider the Kinnersley class I Petrov type D vacuum metric [12] which is a -non-degenerate metric [13]. The Cartan invariants can all be reconstructed from the 4 (complex) scalar polynomial invariants:

where is the Weyl spinor, are spinor indices, and are frame indices. Here, , and are all constants while all of the functions ( etc.) depend only on the coordinate .

Since the invariants only depends on one variable, namely , this spacetime possesses (at least) 3 transitive Killing vectors at any point.

5 Analytic metric continuation

The operators also have another property that seems to be very useful. The metric does not appear explicity in the analysis above and only appears after raising an index: . Consequently, this analysis is independent of the signature of the metric. In the eigenvalue equation the identity operator would therefore be independent of the signature also. We can thus consider what happens under more general coordinate transformations than the real diffeomorphisms preserving the metric signature. Let us therefore consider complex analytic continuations of the real metric of this form.

Consider a point and a neighbourhood, at , and we will assume this nighbourhood is an analytic neighbourhood and that are coordinates on so that . We will adapt the coordinates so that the point is at the origin of this coordinate system. Consider now the complexification of . This complexification enables us to consider the complex analytic neighbourhood of .

Furthermore, let be a complex bilinear form induced by the analytic extension of the metric:

Consider now a real analytic submanifold containing : with coordinates . The imbedding enables us to pull back the complexified metric onto :


In terms of the coordinates : . This bilinear form may or may not be real. However, if the bilinear form is real (and non-degenerate) then we will call it an analytic metric extension of with respect to .

In the following, let us call the analytic metric extension for ; i.e., . We note that this transformation is complex, and we can assume, since is real analytic, that is analytic.

The analytic metric continuation leaves the point stationary. Therefore, it induces a linear transformation, , between the tangent spaces and . The transformation is complex and therefore may change the metric signature; consequently, even if the metric is real, it does not necessarily need to have the same signature as .

Consider now the curvature tensors, and for , and and for . Since both metrics are real, their curvature tensors also have to be real. The analytic metric continuation induces a linear transformation of the tangent spaces; consequently, this would relate the Riemann tensors and through a complex linear transformation. However, how are these related?

By using we can relate the metrics . Since the map is analytic (albeit complex), the curvature tensors are also related via . For the operators this has a very useful consequence. First, we note that scalar polynomial invariants are invariant under ; therefore, the eigenvalues of and are identical at . Consequently, the eigenspace decomposition is identical also. (This seems almost too remarkable to be true). Over the neigbourhoods and we can, in general, relate the eigenvalues: .

Since, a Riemannian space is always characterised by its invariants, we immediately have the following result:

Theorem 5.1.

Assume that a space, , of any signature can be analytically continued, in the sense above, to a Riemannian space (of Euclidean signature). Then the spacetime is characterised by its invariants in either a weak or strong sense.

Note that the result is actually stronger than this because not only is it diagonalisable, if we have a symmetric tensor giving rise to an operator (for example ) then, since the Riemannian space must have a symmetric operator , the eigenvalues are real. Consequently, the eigenvalues at must also be real.

Interestingly, the reverse of the above theorem is just as useful:

Corollary 5.2.

If a spacetime is not characterised by its invariants (weakly or strongly), then there exists no analytical continuation of it to a Riemannian space.

In the above we have restricted to analytic continuations leaving a point fixed. In general, one can consider that does not necessarily leave a point fixed. However, for such complex mappings, we need to be a bit careful with the radius of convergence. On the other hand, assuming convergence, the invariants, , can still be related via . Recall that we are restricting ourselves to neighbourhoods where operators do not change algebraic form.

For spacetime metrics of algebraically special types there are also some useful ’no-go’ theorems for Lorentzian manifolds:5

Corollary 5.3.

If a spacetime is of (proper) Weyl, Ricci or Riemann type N, III or II, then there exists no analytical continuation of it to a Riemannian space.


This follows from the fact that any symmetric operator of a Riemannian space is diagonalisable. Since the Weyl, Ricci and Riemann operator of type N, III and II are not diagonalisable, the corollary follows. ∎

There are also a bundle of similar results that follows from a similar analysis.

Signature and convention

Since analytic metric continuation can change the signature, we may be in a situation in which we are comparing invariants of spacetimes of different signatures. Let us note a word of caution in this regard.

Here, we have (implicitly) assumed that the Riemann tensor is given in a coordinate basis with the standard formula involving the metric and the Christoffel symbols. This formula is invariant under an overall change of signature; for example, is the same as . Consequently, the Ricci tensor, also. However, the sign matters for the Ricci scalar defined as where we get . Therefore, even if the invariants may change with a sign, the overall signature is just a convention. This is also evident from entities constructed from a non-zero curvature “vector” . The norm of this vector is an invariant ; however, in signature this is necessarily positive, ; while signature this is necessarily negative .

It is therefore crucial when comparing invariants through an analytic continuation that we specify the signature. Albeit it is conventional to say that metrics with an overall change of signature are indentical. However, the invariants of such metrics may change by a sign.

Let us be more precise. Consider an overall change of signature: , where . Using a coordinate basis, we see that the Christoffel symbols do not depend on . Consequently, the Riemann tensor, does not depend on either. Moveover, we can also see that the definition of the covariant derivatives, , do not depend on . However, raising or lowering indices introduce . Therefore, depending on whether you raise/lower an odd or even number of indicies to create an invariant, the invariant will depend (if odd) or not depend (if even) on . We can therefore split the invariants into invariants that depend on , denoted , and those that do not, denoted . Since the choice of overall sign is merely a convention we make the following identification:


Note that in the neutral signature case changing the overall sign does not change the signature; consequently, the above map gives a non-trivial equivalence relation between invariants of neutral signature metrics.


Kasner universe:

In 1921 E. Kasner [14] wrote down a Ricci flat metric of Riemannian signature (in dimension 4 but we present it in dimension):


Clearly, this has Killing vectors, , and therefore the invariants can only depend on : . Furthermore, since this is a Riemannian space, this space is characterised by its invariants.

There is a number of analytic metric continuations possible using , where is the imaginary unit and is any of the coordinates except . The so-called Kasner Universe is the Lorentzian version obtained by of signature . Consequently, the invariants of this spacetime are identical to those of Riemannian version of signature . In particular, this means that curvature invariants constructed from the norm of the gradient of an invariant, must be non-negative; i.e., .

Usually, the Lorentzian version is written with signature , which can be obtained from the version; consequently, .

Schwarzschild spacetime:

The 4D Schwarzschild spacetime is given by


There exists an analytic metric continuation (not necessarily unique) to a Euclidean signature given by

This shows that the Schwarzschild spacetime is characterised by its invariants. Furthermore, the invariants are idential to the invariants given by the Euclidean version. Note that there is also a Lorentzian version obtained by with identical invariants. This illustrates the subtleties in the difference of signatures and the invariants.

In 5D the Schwarzschild spacetime is given by


Here, there is a similar continuation, that relates the version with the Euclidean . Interestingly, in 5D we can also do the continuation which turns it into a Euclidean space of signature :

Note that this mapping does not leave a point invariant ( is a singularity and therefore has to be omitted); therefore, the invariants are related via .

A pair of metrics:

Let us consider the pair of metrics:


Now, we can easily see that these can be analytically continued into eachother. This shows that these metrics have indentical invariants. Furthermore, it is also clear that we can analytically continue them into a Euclidean metric of signature . Consequently, these metrics are characterised by their invariants.

This pair of metrics illustrate two things. First, the metric (17) is of Segre type and is consequently (weakly) -non-degenerate6. The metric (18), on the other hand, is actually a Kundt degenerate metric. There are therefore continuous metric deformations of the latter metric to other metrics with identical invariants.

Second, the metric (17) is weakly -non-degenerate but not strongly -non-degenerate. The reason for this is the existence of an analytic continuation between two Lorentzian metrics. We might wonder if such spacetimes (which are weakly but not strongly -non-degenerate) always have an associated analytic continuation.

6 Summary: the 4D Lorentzian case

Let us present a summary of the Lorentzian signature case. All of the results presented here are based on the 4-dimensional (4D) case, but it is believed that they are also true in higher-dimensions.

Let us consider the space of 4D Lorentzian metrics over an open neighbourhood . Consider also, , as the set of scalar polynomial invariants over . The set of invariants can be considered as an element of the Cartesian product of smooth functions over , for some . We can thus consider the calculation of invariants, , from a metric as a map , given by the -map:

This map is clearly not surjective, so for a given , the inverse image , may or may not be empty. Let us henceforth assume that we consider only points in the image, of , and let us consider the sets of metrics with identical invariants; i.e., for a , .

The connected components (in the set of metrics) of this inverse image can be of three kinds:

  1. An isolated point. This would correspond to a metric being -non-degenerate. We will use the symbol for such a point.

  2. A generic Kundt-wedge (or Kundt-tree; it is not entirely clear what the topology of this set is). This corresponds to a set of degenerate Kundt metrics [2], none of which is characterised by its invariants, strongly or weakly. This set would be Kundt metrics connected via metric deformations with identical invariants. We will use the symbol to illustrate a set of this kind.

  3. A special Kundt-wedge containing a metric weakly characterised by its invariants. This set contains degenerate Kundt metrics: however, one of the members of this set is a special Kundt metric which is characterised by its invariants in the weak sense. We will use the symbol, , of a set of this kind. The symbol corresponds to the special point which is (weakly) characterised by its invariants, at the “top” of the Kundt-wedge.

The various sets , will consist of these components. In all of the examples known to the authors, the points are actually connected via analytic continuations. Figure 1 presents a summary of the various cases that are known.

Figure 1: Figures showing the sets of metrics of identical invariants.

Let us consider each figure in turn and comment on the various cases.

  1. The generic case: Strongly -non-degenerate. This case consists of only one isolated point . There is a unique metric with this set of invariants.

  2. Two or more isolated points , all of which are (weakly) -non-degenerate. The Kasner metrics, eq. (14), are examples of this kind.

  3. Two or more components of types and . The pair of metrics, eqs. (17) and (18), are examples of this type and the metrics given corresponds to the two points .

  4. One, two or more components of type . As an example of this kind are the two metrics and .

  5. A generic Kundt-wedge, . Examples of this case are generic degenerate Kundt metrics.

It is not clear whether there might exist sets of type combined with any of the other components (like ) but no such examples are known to date. If consists of several components, then in all of the examples known to the authors, there is always an analytic metric continuation connecting points in them (not necessarily all points of the components).

7 Neutral Signature

Let us also briefly discuss the 4D pseudo-Riemannian case of neutral signature (NS space). Here, little has previously been done with regards to the connection between the invariants and the NS space.

Let us first consider the Segre types of the Ricci tensor. This is essentially the canonical forms of of the Ricci operator and these will also be given for the non-standard types. Note that the canonical forms are obtained using real transformations. The possible types are [15]:

  • : All real eigenvalues. Degenerate cases are: , , , , , , .

  • .

  • .

  • : 4 equal eigenvalues. Canonical forms for Ricci operator, and metric, :

  • : Two non-diagonal Jordan blocks with 2 distinct eigenvalues. Canonical form: