The index of coincidence is useful both in the analysis of natural-language plaintext and in the analysis of ciphertext (cryptanalysis). Even when only ciphertext is available for testing and plaintext letter identities are disguised, coincidences in ciphertext can be caused by coincidences in the underlying plaintext. This technique is used to cryptanalyze the Vigenère cipher, for example. For a repeating-key polyalphabetic cipher arranged into a matrix, the coincidence rate within each column will usually be highest when the width of the matrix is a multiple of the key length, and this fact can be used to determine the key length, which is the first step in cracking the system.
Coincidence counting can help determine when two texts are written in the same language using the same alphabet. (This technique has been used to examine the purported Bible code.) The causal coincidence count for such texts will be distinctly higher than the accidental coincidence count for texts in different languages, or texts using different alphabets, or gibberish texts.
To see why, imagine an "alphabet" of only the two letters A and B. Suppose that in our "language", the letter A is used 75% of the time, and the letter B is used 25% of the time. If two texts in this language are laid side-by-side, then the following pairs can be expected:
Now consider the case when both messages are encrypted using the substitution cipher which replaces A with B and vice versa:
Now suppose that only one message (say, the second) is encrypted using the same substitution cipher (A,B)→(B,A). The following pairs can now be expected:
The same principle applies to real languages like English, because certain letters, like E, occur much more frequently than other letters—a fact which is used in frequency analysis of substitution ciphers. Coincidences involving the letter E, for example, are relatively likely. So when any two English texts are compared, the coincidence count will be higher than when an English text and a foreign-language text are used.
It can easily be imagined that this effect can be subtle. For example, similar languages will have a higher coincidence count than dissimilar languages. Also, it isn't hard to generate random text with a frequency distribution similar to real text, artificially raising the coincidence count. Nevertheless, this technique can be used effectively to identify when two texts are likely to contain meaningful information in the same language using the same alphabet, to discover periods for repeating keys, and to uncover many other kinds of nonrandom phenomena within or among ciphertexts.
The same idea can be applied to a single text, where the sample is in effect compared with itself. Mathematically we can compute the index of coincidence IC for a given letter-frequency distribution as
The products count the number of combinations of elements taken two at a time. (Actually this counts each pair twice; the extra factors of 2 occur in both numerator and denominator of the formula and thus cancel out.) Each of the occurrences of the -th letter matches each of the remaining occurrences of the same letter. There are a total of letter pairs in the entire text, and is the probability of a match for each pair, assuming a uniform random distribution of the characters (the "null model"; see below). Thus, this formula gives the ratio of the total number of coincidences observed to the total number of coincidences that one would expect from the null model.
The expected average value for the I.C. can be computed from the relative letter frequencies of the source language:
If all letters of an alphabet were equally distributed, the expected index would be 1.0. The actual monographic I.C. for telegraphic English text is around 1.73, reflecting the unevenness of natural-language letter distributions. Expected values for various languages are:
|Language||Index of Coincidence|
The above description is only an introduction to use of the Index of Coincidence, which is related to the general concept of correlation. Various forms of Index of Coincidence have been devised; the "delta" I.C. (given by the formula above) in effect measures the autocorrelation of a single distribution, whereas a "kappa" I.C. is used when matching two text strings. Although in some applications constant factors such as and can be ignored, in more general situations there is considerable value in truly indexing each I.C. against the value to be expected for the null hypothesis (usually: no match and a uniform random symbol distribution), so that in every situation the expected value for no correlation is 1.0. Thus, any form of I.C. can be expressed as the ratio of the number of coincidences actually observed to the number of coincidences expected (according to the null model), using the particular test setup.
From the foregoing, it is easy to see that the formula for kappa I.C. is
where is the common aligned length of the two texts A and B, and the bracketed term is defined as 1 if the -th letter of text A matches the -th letter of text B, otherwise 0.
A related concept, the "bulge" of a distribution, measures the discrepancy between the observed I.C. and the null value of 1.0. The number of cipher alphabets used in a polyalphabetic cipher may be estimated by dividing the expected bulge of the delta I.C. by the observed bulge, although in many cases (such as when a repeating key was used) better techniques are available.
As a practical illustration of the use of I.C., suppose that we have intercepted the following ciphertext message:
QPWKA LVRXC QZIKG RBPFA EOMFL JMSDZ VDHXC XJYEB IMTRQ WNMEA IZRVK CVKVL XNEIC FZPZC ZZHKM LVZVZ IZRRQ WDKEC HOSNY XXLSP MYKVQ XJTDC IOMEE XDQVS RXLRL KZHOV(The grouping into five characters is just a telegraphic convention and has nothing to do with actual word lengths.) Suspecting this to be an English plaintext encrypted using a Vigenère cipher with normal A–Z components and a short repeating keyword, we can consider the ciphertext "stacked" into some number of columns, for example seven:
QPWKALV RXCQZIK GRBPFAE OMFLJMS DZVDHXC XJYEBIM TRQWN…If the key size happens to have been the same as the assumed number of columns, then all the letters within a single column will have been enciphered using the same key letter, in effect a simple Caesar cipher applied to a (somewhat "random") selection of English plaintext characters. The corresponding set of ciphertext letters should have a roughness of frequency distribution similar to that of English, although the letter identities have been permuted (shifted by a constant amount corresponding to the key letter). Therefore if we compute the aggregate delta I.C. for all columns ("delta bar"), it should be around 1.73. On the other hand, if we have incorrectly guessed the key size (number of columns), the aggregate delta I.C. should be around 1.00. So we compute the delta I.C. for assumed key sizes from one to ten:
QPWKA LVRXC QZIKG RBPFA EOMFL JMSDZ VDH…We can now try to determine the most likely key letter for each column considered separately, by performing trial Caesar decryption of the entire column for each of the 26 possibilities A–Z for the key letter, and choosing the key letter that produces the highest correlation between the decrypted column letter frequencies and the relative letter frequencies for normal English text. That correlation, which we don't need to worry about normalizing, can be readily computed as
where are the observed column letter frequencies and are the relative letter frequencies for English.
When we try this, the best-fit key letters are reported to be "
EVERY," which we recognize as an actual word, and using that for Vigenère decryption produces the plaintext:
MUST CHANGE MEETING LOCATION FROM BRIDGE TO UNDERPASS SINCE ENEMY AGENTS ARE BELIEVED TO HAVE BEEN ASSIGNED TO WATCH BRIDGE STOP MEETING TIME UNCHANGED XXwhere word divisions have been restored at the obvious positions. "
XX" are evidently "null" characters used to pad out the final group for transmission.
This entire procedure could easily be packaged into an automated algorithm for breaking such ciphers. It should be noted that, due to normal statistical fluctuation, such an algorithm will occasionally make wrong choices, especially when analyzing short ciphertext messages.