diff --git a/README.md b/README.md
new file mode 100644
index 0000000..efc4e53
--- /dev/null
+++ b/README.md
@@ -0,0 +1,106 @@
+# A realtime digital signal processing (DSP) library for Arduino
+
+Filters implements digital filters, which mimic the following common
+analog filters:
+
+ * **FiltersOnePole** implement a digital “RC” type filter, both
+   highpass and lowpass.
+ * **FiltersTwoPole** implements a digital lowpass “RLC” type of filter.
+   Since an RLC filter is a just a dampened harmonic oscillator, the
+   FiltersTwoPole class can be used to implement oscillators as well, by
+   setting the quality factor and resonant frequency to the desired
+   values.
+
+In addition, the Filters library implements a class for keeping track of
+a variable’s statistics in realtime.
+
+ * **RunningStatistics** keeps track variable’s statistics, over a
+   specified window of time (a “running window”), up to the current
+   time. The class calculates variance and mean, as well as the derived
+   values such as standard deviation (sigma) and coefficient of
+   variation.
+
+## Installation
+
+Unzip the code, and place it in the Arduino/libraries directory. (If
+Arduino is already started, it must be restarted to register the
+library.) All of the filter classes can be added to the program by
+declaring
+
+```c++
+#include <Filters.h>
+```
+
+The code can be dowloaded from [The GitHub
+repository](https://github.com/JonHub/Filters).
+
+## Usage
+
+To use a filter, first declare an instance of the class. Various
+parameters, such as the filter frequency, can be set by additional calls
+to the filter class.
+
+The filters are updated by continually calling the `input( newValue )`
+function on the filter, and passing in a new value. For more
+information, see the documentation (below) as well as the code examples
+in the library.
+
+The following code implements a simple one pole (RC type) lowpass
+filter, which can be used for smoothing an analog input. This is
+extremely useful if the input data is noisy, since the filter will
+average together many reads into a single value.
+
+```c++
+// filters out changes faster that 5 Hz.
+float filterFrequency = 5.0;
+
+// create a one pole (RC) lowpass filter
+FilterOnePole lowpassFilter( LOWPASS, filterFrequency );
+
+while( true ) {
+  lowpassFilter.input( analogRead( INPUT_PIN ) );
+  // do something else
+}
+```
+
+## Sampling rates
+
+In typical filter applications, the sampling rate is constant and known
+at the beginning. On the Arduino, however, this is often difficult to
+achieve, as an `analogRead()` within a loop doesn't quite guarantee a
+constant sampling rate. These filters are designed to cope well with
+such situations. The time difference between consecutive samples is
+computed dynamically by taking the difference between the time of the
+current and the previous calls to the filter’s `input()` method. This
+means `input()` should always be called as soon as a data sample is
+available, otherwise the sampling times assumed by the filter will be
+wrong.
+
+Note that handling variable sampling rates comes at a significant cost
+in terms of processing time, which makes these filters a poor choice for
+constant sampling rate applications.
+
+In general, the filter’s output will be more accurate if `input()` is
+called more frequently. Ideally, the delay between samples should be no
+more than the filter’s time constant τ. If the filter is defined by a
+cutoff frequency f<sub>c</sub>, it’s time constant is
+τ&nbsp;=&nbsp;1/(2πf<sub>c</sub>).
+
+## Implementation Notes
+
+The filters are implemented as IIR (infinite impulse response) filters,
+which allows them to operate in real time.
+
+All of the filters remain stable, even if `input()` is not called
+frequently enough. In the case of the single pole (RC type) filters, the
+value of the filter will simply decay to it's resting value. The two
+pole filters use a time-stepped simulation of a damped harmonic
+oscillator. Normally, this type of simulation simulation can become
+unstable if the `input()` is called infrequently.
+
+The code prevents this by setting a maximum allowable timestep
+(approximately a sixth of the resonant period of the oscillator). If the
+`input()` in called less frequently that this, the time step is
+throttled to this value, and the oscillator remains stable.
+
+The filter classes use 32-bit floating point numbers (`float`).