Motivation
A window can be many times brighter than the room around it. A short exposure records the outdoor scene but leaves the room near the noise floor; a long exposure reveals the room but saturates the window. The difficulty is not that the world has run out of colors. A finite measurement range cannot reliably encode every irradiance in the scene under one exposure setting. High dynamic range imaging combines measurements with different useful ranges.
The combined map and its displayed photograph serve different purposes. A radiometric map aims to preserve relative light measurements. A tone-mapped image aims to fit a display and communicate appearance. Increasing local visibility on a display can be helpful without creating any additional measured information. If we mistake a tone-mapped value for radiance, later operations such as illumination or physical compositing can become inconsistent.
This distinction becomes concrete in image-based lighting. A bright environment-map pixel may represent a tiny intense light source, while many moderate pixels represent a broad sky. Their contribution depends on radiance, angular extent and surface orientation. A visually attractive panorama with clipped highlights need not be a useful illumination measurement. We therefore connect acquisition, statistical estimation, display mapping and a simple rendering integral while keeping their assumptions separate.
Definitions
Definition 1.3.1
(Linear exposure measurement and camera response)
For a fixed sensor pixel observing a static scene, let denote sensor irradiance in calibrated units. With exposure duration , an ideal black-level-corrected linear measurement is
The actual sensor also has a saturation limit and a low-signal noise floor. A stored display code may follow a nonlinear response , possibly followed by quantization and clipping. Inverting an ordinary camera response is meaningful only on its invertible unsaturated range. An arbitrary image editing pipeline may not admit one shared scalar response at all.
Definition 1.3.2
(Scene and sensor dynamic range)
The scene's dynamic range is the ratio between largest and smallest relevant positive radiometric values. A sensor's usable dynamic range additionally depends on the saturation threshold and a specified minimum acceptable signal-to-noise criterion. Zero cannot be used as the denominator of a finite ratio. More bit depth supplies finer encoding; it does not automatically lower sensor noise or prevent highlight saturation.
Definition 1.3.3
(Weighted exposure estimator)
For a fixed pixel and a set of valid linear measurements, let and choose deterministic weights with . The weighted estimator is
The word deterministic matters: weights or rejection rules chosen from the observed noisy values can change the estimator's expectation. A practical HDR pipeline needs to handle that selection, alignment and motion in addition to the ideal estimator analyzed here.
Definition 1.3.4
(Global tone curve)
A global tone curve maps a nonnegative scene value to a display value. One simple family first sets for an exposure scale and then uses . Its output lies in . Display encoding is another subsequent transformation, not part of the physical irradiance estimate.
Definition 1.3.5
(Diffuse lighting from an environment map)
For environment lighting, is incoming radiance from unit direction at a surface. A Lambertian surface with diffuse reflectance has outgoing radiance
This expression assumes that the local incident illumination is represented by that environment and visibility is already included or assumed unobstructed. The surface normal has unit length. The integral uses solid angle, not a uniform count of image pixels.
Theorem/proposition
Theorem 1.3.6
(Weighted exposure estimation under independent zero-mean noise)
Assume , independent errors with variances , known positive exposure durations, no clipping, and fixed weights as above. Then is unbiased and
Among unbiased linear combinations of the , the smallest variance is obtained by normalized weights proportional to , and equals .
Theorem 1.3.7
(Constant environment illumination has the correct diffuse scale)
If in every direction, the surface is unoccluded, and the Lambertian model above applies, then for every unit normal. In particular,
This identity is a normalization check for numerical environment-lighting code, not a statement that all real surfaces are Lambertian.
The estimator theorem says how to combine measurements under a specified model. It does not say that longer exposure is always preferable: saturation invalidates the linear equation, motion violates the shared scene value, and sensor variance can depend on exposure and signal. The lighting theorem similarly isolates one material model. Specular surfaces require a directional reflection function and cannot be described by the same diffuse formula alone.
Proof sketch or proof idea
Proof
(Unbiasedness, variance and the optimal weights)
Each normalized measurement satisfies , so its expectation is . A normalized weighted sum remains unbiased. Independence makes the covariance cross-terms zero, giving the variance formula. For arbitrary normalized coefficients with , set . Cauchy–Schwarz gives
Equality requires proportional to , hence . These optimal coefficients are nonnegative. Correlated errors would introduce covariance terms and change the optimal combination; repeated captures do not automatically provide independent information.
Proof X-ray
Which operations keep the estimation claim true?
The unbiasedness proof needs a common unknown , linear unsaturated measurements, and weights fixed independently of measurement noise. Dividing by exposure corrects a known multiplicative change. Clipping is many-to-one and cannot be undone by division. Rejecting a sample because its noisy value exceeds a threshold conditions the remaining noise distribution. The theorem is therefore a baseline for a carefully designed estimator, not a blanket guarantee for all exposure-fusion algorithms.
Proof
(Normalizing the cosine integral)
For diffuse normalization, rotate coordinates so that is the north pole. On the visible hemisphere, and . Therefore
Multiplying by yields the result. The factor is part of solid angle. Omitting it overweights the poles in a latitude–longitude map, even when every stored pixel has equal rectangular area.
An unknown response can be expressed on a log scale as , where and . Adding a constant to both and every leaves these equations unchanged. A normalization, such as fixing one response value, removes the algebraic gauge but does not supply absolute physical calibration. Recovering response and irradiance from limited observations also needs adequate exposure overlap and suitable constraints; a scale convention alone does not establish uniqueness for an arbitrary response function.
Worked examples
Worked example 1.3.8
(Exposure normalization and an unusable highlight)
For durations and in chosen time units, corrected measurements and give irradiance estimates and . If both measurement variances are , optimal weights have ratio , giving . The predicted standard deviation is . If the sensor instead clips at , the recorded second value is not a linear measurement of the same form. Treating as exact irradiance would hide lost highlight information rather than recover it.
Worked example 1.3.9
(Tone mapping changes the meaning of averages)
For scene values and , the radiometric average is five and . Averaging the separately mapped values gives . These differ substantially. Since , the curve is concave and systematically changes averaging. Neither displayed value is an additional physical measurement. If the intended operation is a light mixture or motion blur under a linear model, combine the radiometric values first and tone-map the result afterwards.
Worked example 1.3.10
(A small source can dominate illumination)
Approximate an environment by two angular cells. A cell of radiance occupies steradians with cosine factor one; a broad cell of radiance one occupies one steradian with cosine factor . Their irradiance contributions are approximately and . The tiny bright source contributes twice as much despite its small image area. For , these two cells alone give . Clipping the first radiance to one changes its contribution to and greatly underestimates illumination. The calculation is a two-cell quadrature illustration, not an exact integral for finite cells with varying cosine.
A discretized latitude–longitude environment uses cell solid angles approximately at cell centers, or exact band areas . Multiplying radiance by solid angle and clamped cosine gives a diffuse quadrature rule. Refinement improves the approximation when the angular signal and sampling resolve relevant sources; an unresolved sun can still be missed. Rotating the environment changes its relation to the surface normal and must rotate directions, not merely reinterpret pixel indices.
Concept lensAnalytic
Tone curves trade dynamic range for relative contrast
The curve is strictly increasing because , so it preserves ordering. But its logarithmic slope is for . Bright-region relative contrast is compressed more strongly. Monotonicity does not mean contrast preservation. A local tone mapper may additionally use neighborhood statistics, which can enhance detail but may introduce halos or change the ordering of different regions.
Common mistakes
Counterexample mode
More stored levels cannot restore a saturated measurement
Suppose a sensor reports . With , both and produce one. Re-encoding that value in 16-bit or floating-point storage leaves the two scenes indistinguishable. An additional shorter exposure can separate them if its measurement is unsaturated, but increasing the file's precision cannot invert a many-to-one acquisition step.
Common mistake
(Averaging codes instead of radiance)
Do not average nonlinear camera codes and call the result average radiance, and do not divide a response-encoded code by exposure before it has been linearized. The same warning applies to tone-mapped values: in the example above, averaging the mapped values gives , whereas the mapped radiometric average is .
Common mistake
(Clipping channels after a luminance curve)
Multiplying all three channels by one positive factor preserves their ratios, but clipping individual channels afterwards does not. The color clipped at one becomes , and its red-to-green ratio changes from to . A scalar luminance argument therefore does not by itself guarantee hue preservation on a real display.
Common mistake
(Repeated exposures of a moving scene are not repeated observations)
Motion is also an estimation problem, not merely an aesthetic nuisance. A moving leaf may contribute different surfaces at the same pixel in successive exposures. Perfect knowledge of exposure durations does not turn those measurements into repeated observations of one . Registration, deghosting or a selected reference capture introduces additional modeling choices that should be assessed separately from the variance calculation.
Summary
HDR acquisition combines complementary measurements after accounting for exposure and camera response. The simplest linear estimator has an explicit unbiasedness and variance guarantee under stated noise and validity conditions. Tone mapping changes radiometric relationships to fit a display and should remain conceptually separate from reconstruction. Image-based lighting uses radiance weighted by solid angle and orientation, making unclipped highlights and correct angular sampling mathematically consequential.
Exercises
Quick check
(Two unsaturated independent measurements use durations one and two, both with noise variance . Find the optimal normalized weights on their irradiance estimates and the resulting variance. Compare equal weights.)
Solution · Answer
The normalized estimates have variances and , so inverse-variance weights are . The resulting variance is . Equal weights give , which is larger. This comparison holds for the stated equal measurement variances, rather than assuming that every real sensor has exposure-independent noise.
Quick check
(In , verify the common-additive-constant ambiguity. If is shifted by , how do the recovered irradiances change? Does a normalization alone yield physical units?)
Solution · Answer
Replacing by preserves every equation. With , exponentiating gives . A chosen response normalization fixes one arbitrary scale representative; physical units need an external calibration relating that representative to a measured irradiance or radiance.
Quick check
(Compute and . Compare their display ratio with the scene ratio. Prove directly that preserves order for nonnegative inputs.)
Solution · Answer
The values are and , with ratio instead of three. For , subtraction gives . Thus strict order is preserved, while ratios and contrast generally are not.
Quick check
(A constant environment has radiance two and diffuse reflectance . Compute outgoing radiance. A numerical rule instead returns an unweighted sum that quadruples when every environment pixel is subdivided into four equal-radiance pixels. Explain why that rule cannot approximate the stated lighting integral consistently.)
Solution · Answer
The constant-environment theorem gives . Subdivision does not change the physical environment, so a consistent quadrature must account for the smaller solid angle of each cell. An unweighted sum overcounts samples; doubling, quadrupling or any other growth with mere reparameterization reveals missing integration weights. The cosine factor is also necessary and cannot be replaced by a global sample count.