Comparison of Unicode encodings
This article compares Unicode encodings. Two situations are considered: 8-bit-clean environments, and environments that forbid use of byte values that have the high bit set. Originally such prohibitions were to allow for links that used only seven data bits, but they remain in the standards and so software must generate messages that comply with the restrictions. Standard Compression Scheme for Unicode and Binary Ordered Compression for Unicode are excluded from the comparison tables because it is difficult to simply quantify their size.
A UTF-8 file that contains only ASCII characters is identical to an ASCII file. Legacy programs can generally handle UTF-8 encoded files, even if they contain non-ASCII characters. For instance, the C printf function can print a UTF-8 string, as it only looks for the ASCII '%' character to define a formatting string, and prints all other bytes unchanged, thus non-ASCII characters will be output unchanged.
UTF-16 and UTF-32 are incompatible with ASCII files, and thus require Unicode-aware programs to display, print and manipulate them, even if the file is known to contain only characters in the ASCII subset. Because they contain many zero bytes, the strings cannot be manipulated by normal null-terminated string handling for even simple operations such as copy.
Therefore, even on most UTF-16 systems such as Windows and Java, UTF-16 text files are not common; older 8-bit encodings such as ASCII or ISO-8859-1 are still used, forgoing Unicode support; or UTF-8 is used for Unicode. One rare counter-example is the "strings" file used by Mac OS X (10.3 and later) applications for lookup of internationalized versions of messages which defaults to UTF-16, with "files encoded using UTF-8 ... not guaranteed to work."
UTF-8 requires 8, 16, 24 or 32 bits (one to four octets) to encode a Unicode character, UTF-16 requires either 16 or 32 bits to encode a character, and UTF-32 always requires 32 bits to encode a character. The first 128 Unicode code points, U+0000 to U+007F, used for the C0 Controls and Basic Latin characters and which correspond one-to-one to their ASCII-code equivalents, are encoded using 8 bits in UTF-8, 16 bits in UTF-16, and 32 bits in UTF-32. The next 1,920 characters, U+0080 to U+07FF (encompassing the remainder of almost all Latin-script alphabets, and also Greek, Cyrillic, Coptic, Armenian, Hebrew, Arabic, Syriac, Tāna and N'Ko), require 16 bits to encode in both UTF-8 and UTF-16, and 32 bits in UTF-32. For U+0800 to U+FFFF, i.e. the remainder of the characters in the Basic Multilingual Plane (BMP, plane 0, U+0000 to U+FFFF), which encompasses the rest of the characters of most of the world's living languages, UTF-8 needs 24 bits to encode a character, while UTF-16 needs 16 bits and UTF-32 needs 32. Code points U+010000 to U+10FFFF, which represent characters in the supplementary planes (planes 1-16), require 32 bits in UTF-8, UTF-16 and UTF-32. All printable characters in UTF-EBCDIC use at least as many bytes as in UTF-8, and most use more, due to a decision made to allow encoding the C1 control codes as single bytes. For seven-bit environments, UTF-7 is more space efficient than the combination of other Unicode encodings with quoted-printable or base64 for almost all types of text (see "Seven-bit environments" below).
Each format has its own set of advantages and disadvantages with respect to storage efficiency (and thus also of transmission time) and processing efficiency. Storage efficiency is subject to the location within the Unicode code space in which any given text's characters are predominantly from. Since Unicode code space blocks are organized by character set (i.e. alphabet/script), storage efficiency of any given text effectively depends on the alphabet/script used for that text. So, for example, UTF-8 needs one less byte per character (8 versus 16 bits) than UTF-16 for the 128 code points between U+0000 and U+007F, but needs one more byte per character (24 versus 16 bits) for the 63,488 code points between U+0800 and U+FFFF. Therefore, if there are more characters in the range U+0000 to U+007F than there are in the range U+0800 to U+FFFF then UTF-8 is more efficient, while if there are fewer then UTF-16 is more efficient. If the counts are equal then they are exactly the same size. A surprising result is that real-world documents written in languages that use characters only in the high range are still often shorter in UTF-8, due to the extensive use of spaces, digits, newlines, html markup, and embedded English words.
As far as processing time is concerned, text with variable-length encoding such as UTF-8 or UTF-16 is harder to process if there is a need to find the individual code units, as opposed to working with sequences of code units. Searching is unaffected by whether the characters are variable sized, since a search for a sequence of code units does not care about the divisions (it does require that the encoding be self-synchronizing, which both UTF-8 and UTF-16 are). A common misconception is that there is a need to "find the nth character" and that this requires a fixed-length encoding; however, in real use the number n is only derived from examining the n− 1 characters, thus sequential access is needed anyway.UTF-16BE and UTF-32BE are big-endian, UTF-16LE and UTF-32LE are little-endian. When character sequences in one endian order are loaded onto a machine with a different endian order, the characters need to be converted before they can be processed efficiently, unless data is processed with a byte granularity (as required for UTF-8). Accordingly, the issue at hand is more pertinent to the protocol and communication than to a computational difficulty.
For processing, a format should be easy to search, truncate, and generally process safely. All normal Unicode encodings use some form of fixed size code unit. Depending on the format and the code point to be encoded, one or more of these code units will represent a Unicode code point. To allow easy searching and truncation, a sequence must not occur within a longer sequence or across the boundary of two other sequences. UTF-8, UTF-16, UTF-32 and UTF-EBCDIC have these important properties but UTF-7 and GB 18030 do not.
Fixed-size characters can be helpful, but even if there is a fixed byte count per code point (as in UTF-32), there is not a fixed byte count per displayed character due to combining characters. Considering these incompatibilities and other quirks among different encoding schemes, handling unicode data with the same (or compatible) protocol throughout and across the interfaces (e.g. using an API/library, handling unicode characters in client/server model, etc) can in general simplify the whole pipeline while eliminating a potential source of bugs at the same time.
UTF-16 is popular because many APIs date to the time when Unicode was 16-bit fixed width. However, using UTF-16 makes characters outside the Basic Multilingual Plane a special case which increases the risk of oversights related to their handling. That said, programs that mishandle surrogate pairs probably also have problems with combining sequences, so using UTF-32 is unlikely to solve the more general problem of poor handling of multi-code-unit characters.
If any stored data is in UTF-8 (such as file contents or names), it is very difficult to write a system that uses UTF-16 or UTF-32 as an API. This is due to the oft-overlooked fact that the byte array used by UTF-8 can physically contain invalid sequences. For instance, it is impossible to fix an invalid UTF-8 filename using a UTF-16 API, as no possible UTF-16 string will translate to that invalid filename. The opposite is not true: it is trivial to translate invalid UTF-16 to a unique (though technically invalid) UTF-8 string, so a UTF-8 API can control both UTF-8 and UTF-16 files and names, making UTF-8 preferred in any such mixed environment. An unfortunate but far more common workaround used by UTF-16 systems is to interpret the UTF-8 as some other encoding such as CP-1252 and ignore the mojibake for any non-ASCII data.
For communication and storageEdit
UTF-16 and UTF-32 do not have endianness defined, so a byte order must be selected when receiving them over a byte-oriented network or reading them from a byte-oriented storage. This may be achieved by using a byte-order mark at the start of the text or assuming big-endian (RFC 2781). UTF-8, UTF-16BE, UTF-32BE, UTF-16LE and UTF-32LE are standardised on a single byte order and do not have this problem.
If the byte stream is subject to corruption then some encodings recover better than others. UTF-8 and UTF-EBCDIC are best in this regard as they can always resynchronize at the start of the next code point; GB 18030 is unable to recover after a corrupt or missing byte until the next ASCII non-number. UTF-16 and UTF-32 will handle corrupt (altered) bytes by resynchronizing on the next good code point, but an odd number of lost or spurious byte (octet)s will garble all following text.
The tables below list the number of bytes per code point for different Unicode ranges. Any additional comments needed are included in the table. The figures assume that overheads at the start and end of the block of text are negligible.
N.B. The tables below list numbers of bytes per code point, not per user visible "character" (or "grapheme cluster"). It can take multiple code points to describe a single grapheme cluster, so even in UTF-32, care must be taken when splitting or concatenating strings.
|Code range (hexadecimal)||UTF-8||UTF-16||UTF-32||UTF-EBCDIC||GB 18030|
|000000 – 00007F||1||2||4||1||1|
|000080 – 00009F||2||2 for characters inherited from|
GB 2312/GBK (e.g. most
Chinese characters) 4 for
|0000A0 – 0003FF||2|
|000400 – 0007FF||3|
|000800 – 003FFF||3|
|004000 – 00FFFF||4|
|010000 – 03FFFF||4||4||4|
|040000 – 10FFFF||5|
This table may not cover every special case and so should be used for estimation and comparison only. To accurately determine the size of text in an encoding, see the actual specifications.
|Code range (hexadecimal)||UTF-7||UTF-8 quoted-
|UTF-8 base64||UTF-16 q.-p.||UTF-16 base64||GB 18030 q.-p.||GB 18030 base64|
(except U+003D "=")
|1 for "direct characters" (depends on the encoder setting for some code points), 2 for U+002B "+", otherwise same as for 000080 – 00FFFF||1||1 1⁄3||4||2 2⁄3||1||1 1⁄3|
|00003D (equals sign)||3||6||3|
000000 – 00001F
|as above, depending on directness||1 or 3 depending on directness||1 or 3 depending on directness|
|000080 – 0007FF||5 for an isolated case inside a run of single byte characters. For runs 2 2⁄3 per character plus padding to make it a whole number of bytes plus two to start and finish the run||6||2 2⁄3||2–6 depending on if the byte values need to be escaped||4–6 for characters inherited from GB2312/GBK (e.g.
most Chinese characters) 8 for everything else.
|2 2⁄3 for characters inherited from GB2312/GBK (e.g.|
most Chinese characters) 5 1⁄3 for everything else.
|000800 – 00FFFF||9||4|
|010000 – 10FFFF||8 for isolated case, 5 1⁄3 per character plus padding to integer plus 2 for a run||12||5 1⁄3||8–12 depending on if the low bytes of the surrogates need to be escaped.||5 1⁄3||8||5 1⁄3|
Endianness does not affect sizes (UTF-16BE and UTF-32BE have the same size as UTF-16LE and UTF-32LE, respectively). The use of UTF-32 under quoted-printable is highly impractical, but if implemented, will result in 8–12 bytes per code point (about 10 bytes in average), namely for BMP, each code point will occupy exactly 6 bytes more than the same code in quoted-printable/UTF-16. Base64/UTF-32 gets 5 1⁄3 bytes for any code point.
An ASCII control character under quoted-printable or UTF-7 may be represented either directly or encoded (escaped). The need to escape a given control character depends on many circumstances, but newlines in text data are usually coded directly.
BOCU-1 and SCSU are two ways to compress Unicode data. Their encoding relies on how frequently the text is used. Most runs of text use the same script; for example, Latin, Cyrillic, Greek and so on. This normal use allows many runs of text to compress down to about 1 byte per code point. These stateful encodings make it more difficult to randomly access text at any position of a string.
These two compression schemes are not as efficient as other compression schemes, like zip or bzip2. Those general-purpose compression schemes can compress longer runs of bytes to just a few bytes. The SCSU and BOCU-1 compression schemes will not compress more than the theoretical 25% of text encoded as UTF-8, UTF-16 or UTF-32. Other general-purpose compression schemes can easily compress to 10% of original text size. The general purpose schemes require more complicated algorithms and longer chunks of text for a good compression ratio.
Unicode Technical Note #14 contains a more detailed comparison of compression schemes.
Historical: UTF-5 and UTF-6Edit
Proposals have been made for a UTF-5 and UTF-6 for the internationalization of domain names (IDN). The UTF-5 proposal used a base 32 encoding, where Punycode is (among other things, and not exactly) a base 36 encoding. The name UTF-5 for a code unit of 5 bits is explained by the equation 25 = 32. The UTF-6 proposal added a running length encoding to UTF-5, here 6 simply stands for UTF-5 plus 1. The IETF IDN WG later adopted the more efficient Punycode for this purpose.
Not being seriously pursuedEdit
UTF-1 never gained serious acceptance. UTF-8 is much more frequently used.
- Apple Developer Connection: Internationalization Programming Topics: Strings Files
- "Character Encoding in Entities". Extensible Markup Language (XML) 1.0 (Fifth Edition). W3C. 2008.
- Seng, James, UTF-5, a transformation format of Unicode and ISO 10646, 28 January 2000
- Welter, Mark; Spolarich, Brian W. (2000-11-16). "UTF-6 - Yet Another ASCII-Compatible Encoding for ID". Internet Engineering Task Force. Archived from the original on 2016-05-23. Retrieved 2016-04-09.
- Historical IETF IDN WG page