Given a vector: find N orthogonal vectors.

When writing eigen solvers, it’s useful to be able to generate an orthonormal basis containing a particular vector. Another way of saying this is: given some fixed vector, find other vectors that are perpendicular to it and each other. I am under the impression from some websites that this is also useful in computer graphics (but I know nothing about that – let me know if you do!).

I’ve seen people ask how to do this on the internet (for example: here and here) and in general the recommended process is:

• If only one perpendicular vector is required, use a heuristic to come up with something perpendicular (see previous links).
• If two perpendicular vectors are required, create one perpendicular vector using the above heuristic then use the cross-product to come up with the final vector.
• For more than two perpendicular vectors, generate random vectors and use Gram Schmidt to (hopefully) orthogonalise them.
• For 3D graphics, see the “Update” heading below…

The first two usually require branching in the heuristic. The second may not work (if you’re really really unlucky) and requires multiple normalisation operations. Follows is an approach which I haven’t seen documented elsewhere which requires one square root for an arbitrary dimension vector set.

What we want is some technique which creates an orthogonal matrix from a vector. This is exactly what the Householder transformation matrix does:

Where is some unit vector. The matrix is hermitian and unitary and is hence orthonormal. If we can find a simple expression for which will force the first column of to be equal to some unit vector , we have all the information we need to create a basis containing . This turns out to be trivial:

Explicitly, this looks like:

If is real (i.e. ), we set:

Here is a very inefficient C implementation (the square root is not actually necessary – this can be seen in the fact that two entries are always multiplied together and the term in the square root is constant – also the most inner loop over j should start from 1 because when j==0, the value is simply ):

/* The first n values in m is a normalised vector which is desired to be * part of the basis set. Following this vector, n-1 new vectors of length n * will be added which are orthogonal to it. */ void house_ortho(double *m, unsigned n) { const double x = sqrt(1.0 - m[0]); /* As m[0] approaches 1, x approaches 0. This detects that and will end * up forcing the result to be the identity matrix. */ const double s = (x < 1e-20) ? 0.0 : (-1.0 / x); unsigned i, j; /* We first convert the m vector into the v vector. Note that we have * incorporated the multiply by 2 present in the Householder operation * into v by multiplying every element by sqrt(2). */ for (m[0] = x, i = 1; i < n; i++) m[i] *= s; /* Create all of the orthogonal vectors using the definition of the * Householder transformation matrix: I - 2 v v*. */ for (i = 1; i < n; i++) { for (j = 0; j < n; j++) { m[i*n+j] = ((i == j) ? 1.0 : 0.0) - m[i] * m[j]; } } /* Convert the v vector back into m. Probably could have written this * code in a way where this didn't need to happen. */ for (m[0] = 1.0 - x * x, j = 1; j < n; j++) m[j] = -m[j] * x; }

The termination condition for the for loop (over i) in the middle of the code sample could be modified to generate a smaller number of vectors.

This process could be modified to work with a complex valued but is slightly more complicated because is hermitian (which implies that all values on the diagonal are real). The modification amounts to finding a unit vector which forces to be real, multiplying by this unit vector then dividing all resultant vectors by the factor. I haven’t tested this, but it should work.

Update

I posted a link to this blog on Twitter and reached out to see if anyone else had used this method. Warren Moore helpfully responded and pointed me at this paper “Building an Orthonormal Basis, Revisited” (Pixar) which describes numerical and implementation improvements to an algorithm proposed in this paper “Building an Orthonormal Basis from a 3D Unit Vector Without Normalization” (Frisvad, J. R.).

The algorithm which I’ve described in terms of finding the vector which creates the Householder matrix with the first column equal to some unit vector , ends up producing exactly the same equations (given a vector of dimension 3) given in Frisvad’s paper up to a few sign changes – albeit I’ve gone about defining the problem in a slightly different way. We can demonstrate the equivalence by expanding out the resulting householder matrix given the vector we derived earlier:

Note: we don’t actually need because the householder matrix is symmetric. The matrix can be expressed purely in terms of as:

The Pixar paper addresses the singularity which occurs as tends towards 1. We can do the same thing here by negating the definition of the Householder matrix which still produces a matrix with the same properties. i.e.

In this case, the vector is now defined as:

Which moves the singularity to occur as tends towards -1. The matrix still looks very similar:

Efficient implementation of 3-dimension basis finder

As shown in the Pixar paper, there is no need to test the sign of and pick a code path to follow (which would create a branch prediction performance hit) – we can simply extract the sign and use it as a multiplier to flip signs where they need to be flipped (if they need to be flipped).

Modifying some signs in the explicit matrices given previously, we can re-write the matrices as:

Or more simply:

Where:

And here’s the C implementation:

/* The first three values in v are a three point unit vector. The * function produces two more three point vectors of the same coordinate * space which are orthogonal. */ void ortho3(float *v) { float q1 = v[0]; float q2 = v[1]; float q3 = v[2]; float sign = copysignf(1.0f, q1); float scale = 1.0f / (q1 + sign); float q2scale = q2 * scale; float q3scale = q3 * scale; float q3q2scale = q2scale * q3; v[3] = q2; v[4] = q2 * q2scale - sign; v[5] = q3q2scale; v[6] = q3; v[7] = q3q2scale; v[8] = q3 * q3scale - sign; }

Three step digital Butterworth design.

This guide omits a fair amount of detail in order to provide a fairly quick guide.

Butterworth filters have their poles evenly spaced on a circlular arc in the left hand plane of the [s/Laplace/analogue]-domain.

For a low-pass filter of order , the complex valued pole locations are:

Where and is the cutoff frequency in radians.

The transfer function in the s domain is:

Because the poles always occur in conjugate pairs for this kind of filter, it can always be expanded out into a polynomial with real-valued coefficients :

2) Convert the s-domain transfer function to the z-domain.

There are a few methods to convert an s-domain transfer function this into a [z/digital/discrete]-domain transfer function. The most common is the bilinear transform which makes the substitution:

Where:

Or if you’re performing “frequency warping”:

I won’t go into verbose details here, but you probably want frequency warping. The bilinear transform is a linear approximation to the function which maps the s-domain to the digital z-domain. This linear mapping is analogous to the first term of a series expansion of which is most accurate around z=1 (i.e. DC). The frequency warping operation moves the point where most accuracy is preserved to . For higher frequency cutoffs, this becomes important. can be set to the cutoff frequency of the filter to improve the frequency response near the transition.

For a 3-rd order system, the conversion yields (see notes at the end of this page about risks):

Which can be expanded into a rational polynomial transfer function in .

3) Convert from the z-domain to a difference equation.

First, you must make your z-domain expression causal. Positive powers of must be removed otherwise your filter would need to know the future! This is simple, just modify the numerator and denominator of the transfer function by the inverse of the highest power of . For example:

Multiply the numerator and denominator by .

Get as a function of .

The difference equation can be found by substituting z-powers as offsets to the data they are using. i.e. for the above case:

That is what you want to implement in your code.

Things to keep in mind

High-order systems can start behaving very poorly when designed in this way. Even on floating point, once a system reaches 4-th order or higher, if there are poles close to the real axis in the z-domain, the output can become severely degraded. There are ways to mitigate this: the simplest of which is to break the implementation into separate 2nd order filters (biquads).

Real-time re-sampling and linear interpolation?

Disclaimer: I’ve intentionally tried to keep this post “non-mathy” – I want it to provide a high level overview of what linear interpolation does spectrally and provide some evidence as to why it’s probably not suited in audio processing… unless distortion is desirable.

In the context of constant pitch shifting (the input and output signals have a fixed sampling rate), linear interpolation treats the discrete input signal as continuous by drawing straight lines between the discrete samples. The output signal is constructed by picking values at regular times.

In the above, the green arrows are the input samples and the red arrows are where we want to pick the values off. It’s an intuitive answer to “find the missing values”, but what does it actually do to an audio signal? To find this out, it helps to look at the problem in a different way; we can change the intuitive definition described previously to: the linear interpolator interpolates (meaning: inserts a fixed number of zeroes between each input sample) the input signal by some factor, convolves the response with a triangular shaped filter kernel then decimates by some other factor. This is not quite as trivial as the previous definition, but is identical and we can draw the behaviour and system as:

If you are not familiar with signal processing, and the block diagram in the above picture scares you what you need to know is:

• Audio data is real valued and real valued signals have a symmetric magnitude spectrum about DC (in an audio editor, you will only ever see one side, so you’ll just need to imagine that it has a symmetric reflection going from 0 to -pi (pi can be thought of as the Nyquist frequency of the audio i.e. 24 kHz for a 48 kHz input).
• Interpolators insert U-1 zeroes between each sample. This is analogous to shrinking the spectrum of the input by a factor U and concatenating U-1 copies of it. The copies of the spectrum are called “images”. i.e.

• Decimators drop D-1 samples for every input sample. This is analogous to expanding the spectrum by a factor D and wrapping the result on top of itself (there is also an attenuation by D, but I will not draw that). The parts of the spectrum which have been wrapped back onto itself are called “aliases”. i.e.

In audio, aliasing represents a distortion component which usually sounds dreadful. The only way to avoid the aliasing distortions is to ensure that the input signal is band-limited prior to decimation.

• The H(z) block is a filter, this is a convolution applied to the samples that it sees with some other signal. It is analogous to multiplying the spectrum by some other shape.

So, the interpolation operation introduces images which H(z) needs to remove, and the decimation operation will introduce aliases if we try to decimate too much. Typically, in our real-time re-sampler use, we like to fix the interpolation factor U and permit D to vary (this allows us to use an efficient implementation structure). For H(z) to block the images, we know that it must preserve as much as possible the first 1/U component of the spectrum and must attenuate heavily everything from that point up. Here is the response of H(z) for a linear interpolator based re-sampler with an up sampling factor of 4:

Linear interpolator response for up-sample factor of 4.

This is not good – remember, we wanted the spectrum to preserve as much signal as possible for the first quarter of the spectrum and attenuate everything everywhere else. We can see that the worst case level of an imaging component will be about 6 dB below the signal level. It’s worth mentioning here that the problem does not get any better for higher values of U.

There is nothing stopping us from using a proper low-pass filter for H(z) instead of the triangular shape. Here are a few other options for use as a comparison:

The linear interpolator with two additional FIR filters.

The blue and green responses correspond to 8*U and 12*U length FIR filters respectively. These are both reasonably longer than the linear interpolator which has a filter of length 2*U. The way these filters were designed is outside the scope of this article. The red linear interpolator response costs two multiplies per sample to run, the blue costs eight, the green costs twelve – so these filters show a tradeoff between filter quality and implementation complexity. Note that both the blue and green filters achieve at least 50 dB of aliasing rejection – but we pay for this in the passband performance. If the input were an audio signal sampled at 48 kHz, the frequencies between 0 and 24 kHz would map to the frequency range on the graph between 0 and 0.25 (as we are interpolating by a factor of 4). At 18 kHz, we are attenuating the signal by about 11 dB; at 21 kHz, we are attenuating by about 31 dB. There is an interesting question here as to whether this matters as the frequency is so high. We can get around this to a certain extent by pre-equalising our samples to give a subtle high-frequency boost – but that is an extra complexity in the sampling software. Really the only way to make the cutoff sharper is to use longer filters – and that’s not really an option if performance is important.

Here are some examples comparing the output of the above 8-tap per output sample filter to the linear interpolation filter:

A white noise input being resampled using 8-tap polyphase filters (left) and linear interpolation (right).

The input signal to the above output was white noise. Given that the input signal only had content from 0-24 kHz, we would expect that the output signal would only contain information from 0-12 kHz after halving the playback rate. We can see the linear interpolator has “created” a large amount of data in the high frequency region (all from badly attenuated images and aliasing). The “designed” filter attenuates the aliasing heavily but also attenuates some of the high frequency components of the input noise signal.

A complex tonal input being resampled using 8-tap polyphase filters (left) and linear interpolation (right).

The input signal to the above output was a set of tones separated by octaves in frequency. The aliasing components of the spectrum have introduced inharmonic audible distortions in the linear interpolation case. The “designed” filter almost eliminates the distortion. Magic.

I suppose this all comes down to complexity: two multiplies per output sample for linear interpolation vs. more-than-two for a different filter. I chose 8 and 12 for the taps per polyphase component in the examples on this page as I was able to get implementations of the re-sampler where the two sets of filter states (for stereo samples) were able to be stored completely in SSE registers on x64 – this greatly improves the performance of the FIR delay line operations.

Windows Phone, ownCloud and CardDAV and CalDAV

This is a combination of information from the following two locations:

What is this?

This page gives an outline of how to get your Windows Phone to sync calendars and contacts with an ownCloud instance running on a server using a self-signed certificate. This whole process is a hack and I’m incredibly disappointed that Windows Phone does not support this natively – especially since they have a CalDAV and CardDAV implementation already available (used by iCloud and Google accounts). If this process stops working at some point, that should be expected.

I successfully got this working on a Lumia 735.

Process

I’m assuming that you’ve got ownCloud installed on a server using SSL. When the certificate was set up, the FQDN must really be the FDQN. i.e. if your ownCloud instance is hosted at “a.b.com”, the certificate must be for “a.b.com” – not “b.com”. I had this wrong on my home server and was getting the “80072F0D” error on the phone. If you have not yet set up ownCloud or have not set up SSL for your server yet, there are sites already documenting this process How to Create and Install an Apache Self Signed Certificate.

You then need to get the certificate installed on your phone. If you do not do this, you will get errors when the phone tries to start syncing. You can do this by opening the certificate in Internet Explorer on the phone (you can do this easily by copying the certificate to your ownCloud files, logging into your cloud in IE on the phone and opening the file.

You should then be able to follow the instructions here: Setting up CardDAV and CalDAV on Windows Phone 8.1.