## 1. Introduction

Combining constraints from sediment pore fluid profiles of *δ*^{18}O and [Cl^{−}] with *δ*^{18}O of carbonate of benthic foraminifera shells (*δ*^{18}O_{c}), Adkins et al. (2002) inferred that there were larger-than-modern density differences between deep ocean water masses at the Last Glacial Maximum (LGM), a period lasting from 23 to 19 thousand years before present (ka BP; MARGO Project Members 2009), and that differences in density were due primarily to salinity. Of the ocean water masses considered, Adkins et al. (2002) concluded that Glacial Southern Source Bottom Water (GSSBW), deep water originating from the Southern Hemisphere, was the densest due to its high salinity. These results contrast strikingly with the distribution of today’s deep ocean water masses whose density differences are set primarily by temperature; modern southern source bottom water, Antarctic Bottom Water (AABW), is the densest deep ocean water mass because it is cold, while remaining less saline than overlying water masses [see discussion in Miller et al. (2012)].

Larger density differences between northern source deep water and southern source bottom water imply higher vertical stratification. A greater vertical stratification would help explain indications that there was a physically isolated reservoir of CO_{2} in the deep ocean at the LGM (Broecker and Barker 2007). While the LGM distribution of *δ*^{13}C, Cd/Ca, and *δ*^{18}O possibly indicates a different-than-modern ocean circulation (e.g., Marchal and Curry 2008; Lund et al. 2011), the LGM ocean distributions of these properties may also be consistent with a modern circulation and differences in surface forcing (e.g., Gebbie 2012; Gebbie and Huybers 2006; LeGrand and Wunsch 1995). Improved constraints on ocean temperature and salinity in the past would facilitate the disentanglement of physical changes from chemical and biological signals and better explain why tracer fields in the past ocean varied so strikingly from those of today’s ocean.

The search for reliable proxies of past deep ocean temperature and salinity has proved difficult. In theory, pore fluid reconstructions of past *δ*^{18}O and [Cl^{−}] have a unique advantage over other paleoceanographic proxies of temperature and salinity in that they respond only to physical changes in the ocean. However, so far pore fluid profiles of *δ*^{18}O and [Cl^{−}] have only been used to reconstruct the temperature and salinity at a single point in time: the LGM. To fully understand the changing ocean circulation over the most recent deglaciation, we must know the temporal evolution of temperature and salinity fields. Here we investigate whether pore fluid profiles can constrain the ocean *δ*^{18}O and [Cl^{−}] at other points in time. Simultaneously, we investigate the extent to which they actually constrain the LGM *δ*^{18}O and [Cl^{−}].

In their reconstructions of LGM *δ*^{18}O and [Cl^{−}], Lado Insua et al. (2014), Malone et al. (2004), Adkins et al. (2002), Schrag et al. (2002), Adkins and Schrag (2001), Paul et al. (2001), Schrag et al. (1996), Schrag and DePaolo (1993), and McDuff (1985) relied on a number of restrictive assumptions that made it impossible to recover histories of *δ*^{18}O and [Cl^{−}] at any time other than the LGM. They supposed that *δ*^{18}O and [Cl^{−}] are both conservative tracers in ocean sediments subject to one-dimensional advection and diffusion.^{1} All of the parameters of the problem were assumed known except the bottom water histories of both *δ*^{18}O and [Cl^{−}]. These histories determine the concentration boundary condition time series at the sediment–water interface. They further assumed that the basic shape of the bottom water histories was known up to a scaling constant; that is, the bottom water histories of *δ*^{18}O and [Cl^{−}] primarily tracked changes in *δ*^{18}O of benthic foraminiferal calcite, benthic *δ*^{18}O_{c}, and sea level, but were able to scale relative to an LGM-to-modern difference in concentration. This scaling parameter was then varied in order to find a good fit between the modeled output and the measurements (Fig. 1). Three different salinity histories resulting from three choices for the LGM-to-modern scaling parameter are shown in Fig. 1a. The modern pore fluid profile data that would result from the application of each of these histories as the bottom water boundary condition are plotted on top of measurements [Ocean Drilling Program (ODP) site 1093] in Fig. 1b. The reported value of the LGM salinity was determined from the LGM-to-modern scaling parameter that yielded pore fluid profiles with the best fit to the measurements. Thus, the deglacial history at each site was fixed to a composite benthic *δ*^{18}O_{c} record and sea level curve, scaled by a spatially variable (site specific) constant.

Today the bulk of the ocean interior is weakly stratified, such that the deep ocean salinity and *δ*^{18}O are everywhere very close to their mean ocean values. However, coupling the local time series to the global mean time series implicitly assumes that changes in total ocean water volume always produce the same local change in properties. Further, this approach requires that all sites coevolve in the same way at all times. Tying bottom water histories at all sites proportionally to the mean sea level curve prohibits them from expressing independent temporal evolution as they approach present day values. Finally, this assumption requires that there is a linear relationship between a site’s concentration and the global mean. Adkins et al. (2002) note that their best fit models generate systematic misfits between all of the data and model-generated profiles in the upper sediment column, implying that the pore fluid profiles contain information about the deglacial evolutions of temperature and salinity that could not be extracted with their methodology.

Previous pore fluid reconstructions of LGM *δ*^{18}O and [Cl^{−}] have all assumed that the dominant diffusion coefficient *L*^{2}/*T*, where *T* = 20 000 years and *L* was the depth of the maximum value of *δ*^{18}O and [Cl^{−}] in the sediments. For a maximum measured salinity at 40 m below sea floor (mbsf), this relation yields ^{−5} cm^{2} s^{−1}. Recent studies have shown that the LGM occurred at different times for different mountain glaciers and ice sheets, reaching their maxima between 33 and 26.5 ka BP and remaining stable until 19–14 ka BP (Clark et al. 2009), which calls into question both the idea of a synchronous LGM in the ocean and the timing of the LGM at any site in the ocean, particularly given that the ocean equilibration time scales may be up to 10 000 years or even longer (Wunsch and Heimbach 2008).

We develop an inverse framework to reconstruct past ocean bottom temperature and salinity from ocean sediment core interstitial water profiles of *δ*^{18}O and [Cl^{−}] (in combination with *δ*^{18}O of carbonate of benthic foraminifera) that allows us to relax many of the assumptions listed above. In particular, we relax the assumption that bottom water histories scale to the sea level curve, allowing us to test the robustness of previous reconstructions and to examine the extent to which we can extract more information than the LGM salinity and *δ*^{18}O from present-day interstitial water profiles. This new approach to the problem allows us to simultaneously reevaluate the information about the LGM yielded by the pore fluid profiles.

There are infinite bottom water histories that could have generated our observations; the inverse problem is nonunique. However, some of the mathematically acceptable histories may be physically implausible. For example, negative salinities in the inverse solution time series are mathematically possible but physically impossible. Therefore, we seek the full distribution of physically acceptable bottom water time series that could have generated our measurements. Our approach relies on a Bayesian Markov chain Monte Carlo (MCMC) method to reconstruct a distribution of bottom water histories that could generate modern pore fluid profiles of *δ*^{18}O and [Cl^{−}]. A Bayesian approach allows us to easily restrict our solution space. The Bayesian approach also explicitly allows us to separate our prior assumptions and uncertainty from information contained in the measurements and to recover a distribution of inverse solutions with varying probabilities. MCMC techniques allow us to consider the fully non-Gaussian and nonlinear problem, in which we invert for physical parameters such as the diffusion coefficient with arbitrary prior distributions, in a manner robust to local minima and maxima of the inverse solution space. The particular algorithm we use is named the Cascading Adaptive Transitional Metropolis in Parallel (CATMIP; Minson et al. 2013), which is one example of the general class of Bayesian MCMC algorithms.

We first describe our methodological approach to the forward and inverse problems and explain how we encapsulate our knowledge about the deep ocean as probability distributions that can be supplied to a Bayesian MCMC sampler. With synthetic examples, we then evaluate what information about past bottom water histories can be recovered from modern pore fluid profiles of *δ*^{18}O and [Cl^{−}]. In a separate manuscript, we use what we have learned from synthetic examples to invert real data and demonstrate the robust and nonrobust oceanographic constraints that can be recovered from existing pore fluid profiles.

For convenience in the text we often refer to [Cl^{−}] as salinity or *S*, using the conversion factor *S* = (19.354/35)[Cl^{−}]; however, all of the data measurements are derived from pore fluid [Cl^{−}]. In our synthetic examples, we consider the evolution of salinity, but our conclusions apply equally to reconstructions of both *δ*^{18}O and *S*, as well as other conservative tracers.

## 2. Problem framework and inverse approach

### a. Forward problem

^{−}] and

*δ*

^{18}O are chemically inert and that there are no significant local horizontal gradients in the sediment properties. The vertical distribution of each individual tracer

*c*then evolves in time according to a one-dimensional advection–diffusion equation:where

*ϕ*is the sediment porosity (water volume/bulk volume),

*θ*is the sediment tortuosity (traveled distance of diffused ion/direct distance between spaces), and

*D*

_{0}is the diffusivity in water of the chemical species of interest. The variable

*D*

_{0}is a function of temperature and ionic strength, which both can be functions of depth and time in ocean sediments. The variability of

*D*

_{0}with temperature is well-predicted by the Stokes–Einstein relation

*η*is the temperature-dependent viscosity of the water and

*T*is the absolute temperature. Thus,

*D*

_{0}=

*rD*

_{0,T=20°C}=

*r*

*r*is a function of temperature. Here we take

*r*, assuming that the vertical gradient of temperature is steady. The variable

*u*is the vertical velocity of the pore water with respect to the sediment–water interface, which is predominantly due to sediment compaction, and

*z*is the depth in the sediment, defined as increasing downward.

*θ*

^{2}is

*θ*

^{2}=

*ϕf*, where

*f*is the nondimensional formation factor:Laboratory measurements suggest that a good approximation for

*f*is

*ϕ*

^{−n}, where an average value of

*n*over various sediments is 1.8 (Berner 1980). Our final equation describing concentration evolution with time and space iswhere

*ϕ*in our model is a smoothed version of the measured porosity in each core. Assuming that compaction ceases at some depth in the sediment

*u*can be derived from the constant sedimentation rate and porosity data. At the depth of no compaction

*H*, the fluid velocity

*u*and the sediment burial velocity

*w*are the same, that is,

*u*

_{H}=

*w*

_{H}and

*w*

_{H}is the constant sedimentation rate, determined by dating the sediment core. The vertical concentration profile is modeled according to Eq. (4), with a concentration profile as the initial condition and with two boundary conditions. The top (ocean–sediment interface, at

*z*= 0 m) boundary condition is the ocean bottom water history overlying the core site that we seek to constrain. At the lower boundary (

*z*= 350 m), we assign a gradient boundary condition. The evolution of Eq. (4) is heavily dominated by diffusion, making it a stiff partial differential equation. We numerically solve the equation with a second-order implicit–explicit (IMEX) method, in which the advection terms are solved explicitly while the diffusion terms are solved implicitly (Ascher et al. 1995). The time step in the solver is 10 years and the spatial resolution is 50 cm. The total integration time is 125 000 years. Over the integration time of 125 000 years and 350 m of sediment, these choices lead to a two-norm combined round-off and truncation error in the solution of

*O*(10

^{−4}) g kg

^{−1}. This error estimate is derived from integration of an analogous one-dimensional advection–diffusion equation with constant coefficients to which there is a known analytical solution.

### b. Inverse problem

*c*(

*z*)|

_{t=mod}, and from this set of observations we infer the past time evolution of water properties at the ocean sediment–water interface,

*c*

_{0}(

*t*). We can write this problem in general aswhere

**m**represents the unknown boundary conditions;

**d**

_{obs}is the measured data [

*c*(

*z*)|

_{t=mod}], the pore fluid profile; and

*G*is the forward model of advection and diffusion in the sediments. When we solve for boundary conditions only, Eq. (5) is a linear equation that can be written in terms of discrete Green’s functions as a matrix–vector multiplication:where the bold, Arial font indicates that

**m**, as the solution for

**m**in both the continuous and discretized framework is an ill-posed problem (Hansen 1998).

We find traditional regularized or damped inverse methods insufficient for solving our particular ill-posed inverse problem. It is difficult to bound the bias in the inverse solution because of the choice of regularization. Further, inversion for the diffusion coefficient

### c. Bayes’ theorem

**m**is a vector of input parameters and

**d**

_{obs}denotes the observations. Vector

**m**is the set of unknowns we wish to solve for, in our case the ocean bottom water history of salinity or

*δ*

^{18}O and the unknown diffusion parameter

**d**

_{obs}is the sediment pore fluid depth profile of measured concentrations. The probability that a given set of input parameters is correct, given the measured data, is proportional to the product of the prior probability of those input parameters and the likelihood of the data produced by the model using those input parameters. The prior probability

*P*(

**m**) encompasses our knowledge of the input parameters independent of the measurements

**d**

_{obs}. The likelihood

*P*(

**d**

_{obs}|

**m**) is computed using the differences between the measured data and output of the forward model using a given set of input parameters

**m**. The posterior probability distribution, or posterior, is

*P*(

**m**|

**d**

_{obs}).

### d. Bayesian estimation for a linear inverse problem with Gaussian prior parameters

*P*(

**m**|

**d**

_{obs}), equal towhereThe vector

**d**

_{obs}denotes the observations,

_{D}is the error covariance matrix for the data in

**d**

_{obs}, and

_{M}is the prior error covariance matrix for the parameters in

**m**(Tarantola 2005).

### e. Bayesian estimation for nonlinear problem or non-Gaussian prior parameters: Bayesian MCMC parameter estimation

For more general cases, either nonlinear problems or when we consider parameters with arbitrary prior probability distributions, we could determine the posterior probability distribution with a grid-search approach. In such an approach, we would construct an *n*-dimensional grid of parameter values (*n* is the dimension of **m**, the number of unknown parameters), with the ranges of each grid dimension determined by the prior probability distribution *P*(**m**) of each parameter. Each point on the grid would represent a vector set of choices for **m**. The initial spacing of the grid would be arbitrarily coarse and we would run the forward model at every grid point and then refine the grid spacing until the estimate of *P*(**m** | **d**_{obs}) converges. Our inverse problem is high dimensional (22–23 parameters), so that a grid-search approach is intractable [because of the curse of dimensionality (Bellman 1957)]. In a Bayesian MCMC approach, instead of directly computing the multidimensional integral of the probability distribution for each parameter, we build an ensemble of inverse solutions whose density in model space is proportional to the Bayesian posterior *P*(**m** | **d**_{obs}). For this search, we use CATMIP (Minson et al. 2013), which behaves efficiently for high-dimensional problems. CATMIP is initialized with the prior probability distribution and evolves from the prior to the posterior in a series of “tempering” or “cooling” steps. At each cooling step the algorithm recovers an intermediate probability distribution between the prior and posterior distributions. For a given set of parameter values **m**, the likelihood *P*(**d**_{obs} | **m**) is computed by evaluating the forward model and computing a misfit cost function between the model prediction and the measured data.

### f. Parameterization of inverse problem

In practice, we must solve for a discrete rather than continuous set of variables. We therefore frame our inverse problem, the solution for a time series from a vertical concentration profile, as a set of discrete parameters. We are primarily interested in recovering the bottom water history, the time series of [Cl^{−}] or *δ*^{18}O at the ocean floor, but its solution is not independent of the other unknowns. As we show later (see section 7), the solution is particularly sensitive to *O*(10) nodes can be resolved by the data [nodes arranged at fixed points in time, determined by examination of the singular value decomposition (SVD) of the discrete Green’s functions generated from the 1D advection–diffusion equation (Miller 2014)]. Higher temporal resolution of the parameterization increases the computational expense of the fully nonlinear inverse problem without a qualitative change in the inverse solution set. To retain accuracy in the forward finite difference solver, we linearly interpolate between the parameters at fixed times to generate the input to the forward model.

### g. Likelihood function

_{D}has only diagonal terms

*σ*

^{2}, where

*σ*is one standard deviation of the measurement error.

We use a Laplace error distribution rather than the commonly used Gaussian error distribution in order to reduce the effect of data outliers on our solutions. Note that in general the form of the Laplace distribution is ∝ *b* = 2*σ*.

## 3. The prior probability of the bottom water history *P*(m)

We represent our prior knowledge of the salinity or *δ*^{18}O bottom water histories as multidimensional Gaussian distributions. The value of salinity or *δ*^{18}O at each fixed time point in the discrete bottom water history is a parameter in our problem. We assign a prior mean and variance to each node. The mean represents our estimate of the global deep ocean mean, while the variance represents our best guess of the global spread of deep ocean salinities or *δ*^{18}O values around the mean at that time. To ensure smoothly varying solutions, we further impose temporal covariance between node values.

### a. Prior information from sea level records

*h*is the mean ocean depth. For consistency with the assumptions made in Adkins et al. (2002), we take the modern mean ocean salinity to be 34.68 g kg

^{−1}and the LGM mean ocean salinity to be

*μ*

_{S}|

_{LGM}=

*h*

_{mod}

*μ*

_{S}|

_{mod}/

*h*

_{LGM}= 35.9 g kg

^{−1}. Our sea level curve is interpolated from a new compilation of published values (Fig. 2, Table 1).

Sea level compilation.

Conversion from total ice mass to the mean *δ*^{18}O_{w} of the ocean requires knowledge of the spatial and temporal variability of *δ*^{18}O of mountain glaciers and continental ice sheets. Estimates of these variations of *δ*^{18}O in the past are relatively unconstrained by measurements, making it impossible to quantify their associated errors. There are many existing estimates of the past contribution of *δ*^{18}O|_{w} (*δ*^{18}O of water) to the *δ*^{18}O|_{c} (*δ*^{18}O of carbonate) of benthic foraminifera at the LGM (e.g., Emiliani 1966; Shackleton 1967; Dansgaard and Tauber 1969; Chappell and Shackleton 1986; Fairbanks 1989; Mix 1987; Schrag et al. 1996). Duplessy et al. (2002) reviews the traditional ice-mass-based estimates of *δ*^{18}O at the LGM and suggests a lower bound of ocean mean *δ*^{18}O|_{w} of 0.85‰ based on the freezing point of seawater and an upper bound based on ice mass approximations of 1.47‰. We take the best current guess of the LGM mean ocean *δ*^{18}O to be the midpoint of the Duplessy et al. (2002) ice mass bounds, 1.16‰. The equation for deep ocean *δ*^{18}O as a function of time (sea level) is similar to that of *μ*_{S}(*t*) [Eq. (11)], with modern mean *δ*^{18}O = 0.

### b. Prior information from modern ocean property spatial spreads

Today, salinities and *δ*^{18}O everywhere in the deep ocean are close to their mean ocean value. Distributions of *δ*^{18}O and salinity in the past are unknown. Lacking other constraints, we might guess that past spatial distributions of deep *δ*^{18}O and salinity around their mean ocean values were the same as their modern spatial distributions. The full ocean modern distributions of salinity and *δ*^{18}O deeper than 2000 m (LeGrande and Schmidt 2006) are not simple Gaussians (Fig. 3); at best they are a mixture of Gaussians. The bimodal nature of the modern distributions comes from the dominance in the modern deep ocean of North Atlantic Deep Water and Antarctic Bottom Water, two water masses with distinct properties. However, the modern ranges of salinity and *δ*^{18}O at a given site are narrow. The bimodality of the global distributions of salinity and *δ*^{18}O are due primarily to spatial rather than temporal (e.g., seasonal, interannual) variability.

In the vicinity of each site on the ocean floor, the modern distribution of salinity and *δ*^{18}O could be described by a Gaussian. However, we wish to avoid imposing the modern ocean state on the glacial ocean state by assuming, for example, that salinity is high in the North Atlantic and low in the Southern Ocean. A less stringent prior assumption is that the feasible end-to-end ranges of salinity and *δ*^{18}O at any site in the past match the modern deep ocean end-to-end ranges. As a conservative estimate for the modern spread of salinity and *δ*^{18}O, we fit a Gaussian probability distribution envelope around each histogram, as shown by the red curves in Fig. 3. Applying this method gives a modern deep ocean standard deviation of salinity of *σ*_{S} = 0.12 g kg^{−1} and a standard deviation of *δ*^{18}O of ^{−1} in salinity, assuming the LGM height and salinity are known. The error propagation formula is *δh* = 10 m and *δ*^{18}O is not directly related to the sea level; however, Duplessy et al. (2002) suggest an error in their estimate for mean *δ*^{18}O_{w} should be 0.2‰. Summing together the modern variance and the estimated error in sea level yields *σ*_{S} = 0.213 g kg^{−1} (^{2} kg^{−2}) and ^{2}].

### c. Prior information about ocean property temporal variations

*T*is a time scale of covariance and

*i*and

*j*are matrix indices denoting discrete values of the parameterized time series. When

*t*

_{i}−

*t*

_{j}=

*T*, the correlation between the time series values at those times is ~0.6, while if

*t*

_{i}−

*t*

_{j}= 2

*T*, the correlation between the values is ~0.1.

*T*represents roughly half of an ocean site’s memory of its previous salinity or

*δ*

^{18}O value. Time scale

*T*is a spatially variable function of ocean mixing, which will also depend on the global distribution of temperature and salinity (through density). Assuming that large-scale steady-state ocean tracer mixing is a balance between advection and eddy diffusion, that is, as in Ferrari and Wunsch (2009),where

**v**is the fluid velocity;

*C*is the tracer concentration field; and

**K**is the (eddy) diffusivity of a tracer, in general a spatially variable tensor. We can scale Eq. (13) to relate the time scale

*T*to an eddy diffusivity

*κ*using a characteristic length scale

*L*. That is,where 2

*T*is the time needed for an ocean site to forget its previous value. Assuming that mixing predominantly proceeds along isopycnals, recent estimates suggest

*κ*is

*O*(100) m

^{2}s

^{−1}(Zika et al. 2009, 2010). For deep ocean sites under the influence of spatially homogeneous deep water masses, we assume

*L*is at least half the meridional extent of the Atlantic Ocean, roughly 8000 km. With these values for

*L*and

*κ*,

*T*is ~10 000 years.

Our scaling argument assumes a steady-state ocean circulation and is simplistic. Paleoceanographic records suggest that the ocean does not remain in steady state over *O*(10^{4}) years (e.g., Charles et al. 1996; Hodell et al. 2003; Lamy et al. 2004; Lynch-Stieglitz 2004; Pahnke and Zahn 2005; Robinson et al. 2005). Using radiocarbon ages to estimate the reservoir age of the deep ocean, some authors conclude that the ocean can reach a new steady state after 2000 years (Duplessy et al. 1991; Skinner and Shackleton 2005), implying that *T* = 1000 years. However, Wunsch and Heimbach (2008) demonstrate that radiocarbon ages can be misleading; the ocean may need up to 10 000 years to reach a new equilibrium in response to regional changes in surface forcing. In our framework this is equivalent to *T* = 5000 years or more.

### d. Accounting for different-than-modern past ocean property spreads

Adkins et al. (2002) conclude that the spread in LGM deep ocean salinity was larger than it is today. Their results demonstrate that the measurements are consistent with a larger-than-modern spread in salinity, but they do not require that the salinity spread was so large. It is possible that the ocean had a wider salinity spread in the past, but we do not have any a priori information on what the magnitude of the spread was.

Bayesian methods allow us to shrink but not expand the distribution on any given parameter. With an assumed Gaussian prior, posterior estimates more than two standard deviations from the mean of the prior are unlikely, as any parameters with 0 prior probability will, by definition, have 0 posterior probability [recall that *P*(**m** | **d**_{obs}) ∝ *P*(**d**_{obs} | **m**)*P*(**m**), so *P*(**m**) = 0 implies that *P*(**m** | **d**_{obs}) = 0]. To allow for the possibility of the LGM salinities estimated by Adkins et al. (2002), we consider the sensitivity of our solution to wider-than-modern prior variances, in particular a prior variance equivalent to that implied by the spread in Adkins et al. (2002), ^{2} kg^{−2}.

### e. Diffusion coefficient prior

We assign a lognormal prior to the diffusion coefficient *D*_{0} for chloride between 1 × 10^{−6} and 20 × 10^{−6} cm^{2} s^{−1}, varying as a function of temperature, pressure, and ionic content (Li and Gregory 1974). The *D*_{0} of *δ*^{18}O is the relative diffusivity of water oxygen isotopes, which mirrors the diffusivity of *O*(1 × 10^{−5}) cm^{2} s^{−1}, also a function of temperature (Wang et al. 1953). In our nonlinear inversions in which we estimate ^{−5} cm^{2} s^{−1} and the standard deviation of the ln

## 4. Synthetic example framework

We consider several synthetic problems to examine the extent to which we can recover bottom water salinity histories from pore fluid profiles. The details of each synthetic example vary, but all follow the same basic approach. We choose hypothetical bottom water salinity histories, which we refer to as *S*_{BW}(*t*). Using a known *S*_{BW}(*t*) as the top (ocean–sediment interface) boundary condition in the forward problem (see section 2a), we generate synthetic data: an ocean sediment pore fluid salinity profile. Random synthetic Gaussian noise of 0.05% is added to the synthetic data to represent measurement uncertainties. We then use CATMIP to invert the noisy synthetic data and recover a posterior distribution of *S*_{BW}(*t*), which we compare to the known *S*_{BW}(*t*) to determine how well we can recover the true solution by inverting the data. The difference between the posterior (inverse) estimate and the true *S*_{BW}(*t*) is summarized by the root-mean-square difference

Summary of experiments.

## 5. Linear problem: Inversion for bottom water salinity history, *S*_{BW}(*t*) only

### a. Inversion of synthetic stretched sea level example with varying prior assumptions

Our first example is modeled after Adkins et al.’s (2002) reconstructed bottom water salinity history in the Southern Ocean (ODP site 1093), the location with highest inferred salinity at the LGM. We suppose that this was the true bottom water history in order to determine what information about such a history would be retained by the modern measured pore fluid profile. To construct *S*_{BW}(*t*), we start with a presumed mean ocean salinity history, *μ*_{S}(*t*) (see section 3a). We multiply *μ*_{S}(*t*) by a constant factor, determined by assuming that the LGM salinity in the Southern Ocean was 37.1 g kg^{−1} [as inferred by Adkins et al. (2002)]. We call this synthetic *S*_{BW}(*t*) as the “stretched sea level” example. The diffusivity parameter ^{−5} cm^{2} s^{−1}, and *r* and *dr*/*dz* are computed from the temperature profile at ODP site 1093 (Gersonde et al. 1999). The variable *u*_{H} is 0.01 cm yr^{−1}, and the model domain (sedimentary column) is 350-m long. The initial salinity profile is interpolated from the measured chlorinity profile (converted to salinity) at ODP site 1093, as in Adkins et al. (2002). Here we assume that the cyclical nature of glaciations and deglaciations would have caused the depth profile of [Cl^{−}] to be the same at the last interglacial (125 000 yr BP) as it is today. The bottom boundary condition is

For the purpose of our present work, we assume that the mean ocean salinity history *μ*_{S}(*t*) is known up to an error of 0.093 g kg^{−1} (see section 3b). Deep ocean salinities should be near the mean ocean value; however, we do not have any prior information on the variance of salinity at the LGM. The range in deep ocean salinities at the LGM reported in Adkins et al. (2002) corresponds to a variance of approximately ^{2} kg^{−2}, while a generous estimate of deep ocean salinity variability from existing modern compiled measurements combined with LGM sea level error is ^{2} kg^{−2}. The full range of modern ocean salinity is encompassed by a variance ^{2} kg^{−2} (one standard deviation range is ~25–45 g kg^{−1}). Similarly, we expect temporal correlation in *S*_{BW}(*t*), but we also have few constraints on the value of the correlation time scale *T*. As discussed in section 3, *T* could be as short as a few hundred years or it could be greater than 5000 years.

With these points in mind, we invert the synthetic data assuming prior variances ^{2} kg^{−2}. As shorthand, we refer to these as “modern,” “intermediate,” and “uninformative” variances, respectively. We also vary the prior assumed covariance time scale *T* and consider its effect on the posterior distribution of *S*_{BW}(*t*).

The results of the uninformative prior variance example, ^{2} kg^{−2}, illustrate that the measured data constrain the absolute values of salinity between 6 ka BP and present when *T* is assumed to be 0 years (Fig. 5). Note that even when *T* is assumed to be 0 years, there is some temporal correlation implicit in *μ*_{S}(*t*). By increasing the value of *T* while holding *S*_{BW}(*t*) is a strong function of *T*. Increasing *T* decreases the posterior variance of *S*_{BW}(*t*), particularly between ~26 ka BP and the present (Fig. 6). However, a reduction of the posterior variance relative to the prior variance of *S*_{BW}(*t*) does not guarantee that the mean posterior estimate aligns with the true solution, as demonstrated in Fig. 5.

As shown in Fig. 5, the mean posterior estimate of *S*_{BW}(*t*) (here equivalent to the mode or highest-probability inverse solution) oscillates around the true synthetic *S*_{BW}(*t*) at times that have a wide posterior variance. This oscillation suggests that while the data do not constrain the value of salinity at these times, they do constrain temporal correlations in *S*_{BW}(*t*), that is, the relative value of salinities at different times in the bottom water history. For this reason we compare the frequency content of the prior and posterior estimates (Fig. 7). The frequency content of the prior *μ*_{S}(*t*) is set equal to the frequency content of the true synthetic stretched sea level *S*_{BW}(*t*), so it is unsurprising that the frequency content of the posterior estimates are similar to the frequency content of the true *S*_{BW}(*t*). Imposing nonzero values of *T* does filter the frequency content of the posterior estimate, removing power at the upper and lower ends of the frequency range. In this particular example, the data do not constrain the frequency content of the bottom water salinity history separately from the prior assumptions.

Assuming ^{2} kg^{−2} and *T* = 0 years in the prior distribution, the synthetic *S*_{BW}(*t*) falls outside one standard deviation of the posterior distribution (Fig. 8). Recall that the synthetic *S*_{BW}(*t*) in this example is inconsistent with the modern prior variance assumption, because it is modeled after the high-salinity Southern Ocean case in Adkins et al. (2002). Increasing the prior variance assumption from 0.05 to 0.5 g^{2} kg^{−2}, holding *T* constant at 0 years, widens the posterior variance but does not significantly alter the mean posterior time series (cf. Fig. 9a to Fig. 8a).

Increasing the prior covariance time scale *T* from 0 to 4000 years improves the match of the posterior mean estimate *S*_{BW}(*t*) to the true synthetic *S*_{BW}(*t*) over the time period between 40 ka BP and present (Figs. 8b, 9b). The improved match of the posterior mean estimate of *S*_{BW}(*t*) to the true synthetic *S*_{BW}(*t*) when *T* = 4000 years results from the underlying temporal correlation of the synthetic *S*_{BW}(*t*) (see Fig. 7). As expected, a more accurate prior guess leads to a more accurate posterior estimate.

### b. Inversion for random time series consistent with the Gaussian prior assumption

In section 5a we found that our ability to recover *S*_{BW}(*t*) from the modern pore fluid profile decreases with time in the past. We also showed that prior assumptions more consistent with the true synthetic *S*_{BW}(*t*) led to more accurate posterior estimates. However, the prior assumptions in section 5a were not strictly consistent with the known *S*_{BW}(*t*). Specifically, some values of the synthetic *S*_{BW}(*t*) were outside the range prescribed by the prior variances. The term *S*_{BW}(*t*) had the same frequency content as the prior *μ*_{S}(*t*), but the prior distribution had an additional constraint of a covariance time scale *T*. As shown in section 5a, the prior constraint *T* reduces the frequency content of the posterior mean estimate relative to the frequency content of the prior mean guess and the true *S*_{BW}(*t*). For this reason, the posterior estimate may be dominated by frequencies or harmonics consistent with the prior covariance time scale, and an accurate estimate of tracer variability at one frequency may trade off with the fit to another, such that parts of the inverse estimate (e.g., millennial variability, centennial variability) are closer to the true synthetic than others.

Because the prior constraints are parameters of a multivariate Gaussian distribution, complete consistency between prior assumptions and a synthetic example requires considering a distribution of synthetic examples. We therefore construct examples in which we randomly draw an ensemble of 10 synthetic *S*_{BW}(*t*) from a multivariable Gaussian distribution consistent with (equivalent to) the prior distribution. In these examples, the mean of the Gaussian is equal to *μ*_{S}(*t*), and we show representative cases for two different variances, ^{2} kg^{−2}, permuted with two different covariance time scales, *T* = 0 and 4000 years. We show that even when the synthetic time series are completely consistent with the prior distribution, the pore fluid data still only constrain salinity a few thousand years before present day.

Even when the prior distribution is consistent with the parent (source) distribution of the synthetic *S*_{BW}(*t*), larger values of *T* and smaller values of *T* = 0 years and ^{2} kg^{−2}, at best the posterior mean estimate resembles the synthetic salinity history in the most recent 8 or 6 ka BP (Fig. 10b). Further back in time, the posterior mean estimate decreasingly resembles the synthetic salinity history, more so the more the synthetic *S*_{BW}(*t*) deviates from the prior mean. In the recent past, if the true synthetic *S*_{BW}(*t*) is very different from the prior mean, it is easier to recover through the inversion. In the past, if the true synthetic *S*_{BW}(*t*) is very different from the prior mean, it is harder to recover through inversion of the measured pore fluid profile data.

The difference between the posterior estimate and an individual random synthetic *S*_{BW}(*t*) varies. However, the ensemble difference between the posterior mean and the synthetic time series is reduced relative to the ensemble difference between the prior mean and the synthetic time series, and the degree of reduction is a strong function of *T* (Fig. 11). With prior assumption *T* = 0 years, the posterior mean estimate for *S*_{BW}(*t*) is more accurate than the prior estimate between 10 ka BP and the present. With *T* = 4000 years, the posterior mean estimate is more accurate than the prior mean estimate between 20 ka BP and present.

## 6. Nonlinear problem: Simultaneous inversion of bottom water salinity and diffusivity

The diffusivity in ocean sediment pore waters ^{−}] can vary up to an order of magnitude from site to site on the ocean floor. Because the value of *S*_{BW}(*t*). We repeat all of the synthetic examples from section 5 in which ^{−5} cm^{2} s^{−1}.

### a. Nonlinear inversion of synthetic stretched sea level example with varying prior assumptions

A larger range of ^{2} kg^{−2} and *T* ranging from 0 to 6000 years, the posterior distributions of *S*_{BW}(*t*) are wider than those of the corresponding linear examples. The true value of

The posterior estimate of *S*_{BW}(*t*) was higher only in the distant past, *T* increases the correlation between *S*_{BW}(*t*).

When we solve for *S*_{BW}(*t*), the frequency content of the posterior mean estimate differs from that of the prior mean and true synthetic *S*_{BW}(*t*). Different values of *S*_{BW}(*t*) (as judged by their frequency content) to create similar pore fluid profile data. Though the minimum and maximum posterior estimates of ^{2} kg^{−2}) and *T* = 0 or 4000 years, *S*_{BW}(*t*) associated with the minimum *S*_{BW}(*t*), and *S*_{BW}(*t*) associated with the maximum *S*_{BW}(*t*). With the intermediate prior variance assumption (^{2} kg^{−2}), the most significant frequencies in the posterior distribution of *S*_{BW}(*t*) break into different harmonic groups at the extreme values of

### b. Nonlinear inversion for random time series consistent with the Gaussian prior assumption

We use the identical ensembles of synthetic *S*_{BW}(*t*) and corresponding synthetic data described in section 5b, but now we solve simultaneously for

In these nonlinear inversions, the posterior mean frequency content is a worse estimate of the true frequency content of the synthetic *S*_{BW}(*t*) than in the linear inversions (Fig. 15). The inverse problem is significantly nonlinear in *S*_{BW}(*t*). In all of these cases the (peak) power in the inverse solution is more broadly distributed at the maximum

As in the stretched sea level examples, the posterior distributions of the estimates of *S*_{BW}(*t*), increased prior variance of *S*_{BW}(*t*) leads to increased variance in the posterior estimate of ^{2} kg^{−2} rather than 0.05 g^{2} kg^{−2}) allows a greater spread in possible

## 7. Discussion and summary

Synthetic examples allow us to determine the maximum information about the past ocean that can be extracted from pore fluid profiles of *δ*^{18}O and [Cl^{−}]. Without assuming a prior covariance time scale or a narrow range of salinity, the measured pore fluid data constrain the bottom water history between 6 ka BP and present. When the assumed range of possible salinity is narrower (^{2} kg^{−2}) the measured pore fluid data constrain the bottom water history between 10 ka BP and present. The difference between the distribution of possible inverse solutions and the true bottom water salinity history is a strong function of the assumed time scale of covariance *T*. Our results demonstrate that modern profiles of pore fluid *δ*^{18}O and [Cl^{−}] can constrain the range of salinity and *δ*^{18}O at the LGM only if *T* is known. When there is no prior assumption of *T*, the spread of possible salinities at the LGM is equivalent to the initial guess or the variance of the prior distribution. While the absolute values of salinities prior to 10 ka BP are unconstrained by the pore fluid data, the data do constrain the correlation between the salinities at different times in the past. This information about correlation is useful because, combined with other independent data, it could narrow the range of past salinity.

Previous work suggested that the diffusivity coefficient

In all the synthetic cases described here, the bottom boundary condition (vertical gradient of *S* at *z* = 350 m) and the initial condition (vertical profile of *S* from the sediment–water interface at *z* = 0 m to *z* = 350 m at 125 000 years ago) are fixed and assumed to be known. We assumed that the cyclical nature of glaciations and deglaciations caused the data profile at 125 ka BP (our initial condition) to be equivalent to the modern measurement profile. We also assumed that the gradient boundary condition at the bottom of the sediment column is constant over the time period of integration and that it can be estimated from the slope of the measured pore fluid data over the final (bottom) tens of meters in the sediment column.

Our inverse framework allows us to relax these assumptions and instead solve for the initial condition and bottom boundary condition as additional unknown parameters. Our preliminary work in this area (see, e.g., Miller 2014) confirms our intuition that, in general, an increased number of free parameters in an ill-posed problem such as this one always increases the uncertainty of the posterior parameter distributions. We make many other assumptions in our advection–diffusion model [Eq. (4)], such as the assumption of steady-state compaction and the assumed relationship between sediment tortuosity and porosity. If these assumptions were relaxed, it is likely that the uncertainty in bottom water salinity and *δ*^{18}O deduced from pore water profiles would further inflate.

Adkins et al. (2002) relied on stretching and shrinking the salinity history scaled to a composite benthic *δ*^{18}O_{c} and mean sea level history, rather than allowing the salinities at each time to vary independently, and this is equivalent to assuming that the salinity covariance time scale *T* is 4000 years or longer. We have shown here that the variance of the inverse estimate of salinity at the LGM (assuming the LGM occurred around 20 ka BP) is entirely determined by the prior assumption of *T*. Therefore, the error reported in Schrag and DePaolo (1993), Schrag et al. (1996), Adkins and Schrag (2001), Paul et al. (2001), Adkins et al. (2002), Schrag et al. (2002), Malone et al. (2004), and Lado Insua et al. (2014) may underestimate the true uncertainty of estimated LGM deep ocean salinity and *δ*^{18}O.

Adkins and Schrag (2003) investigated the sensitivity of estimating LGM salinity to various assumptions about the bottom water salinity history, the diffusivity of [Cl^{−}], the effective advective velocity of [Cl^{−}], and the initial condition pore fluid profile. They concluded that the overall uncertainty in the relative difference between LGM and present salinity was 0.1%–0.5% (out of 3.5%). Their approach involved small perturbations around a fixed value, varying only one parameter at a time. Our approach allows us to vary the parameters simultaneously and therefore explore the inverse solution space more completely. We find that salinities, diffusivity, and initial conditions can compensate for each other. This ability to compensate increases the uncertainty on any one parameter.

Previous authors (Lado Insua et al. 2014; Malone et al. 2004; Adkins et al. 2002; Schrag et al. 2002; Adkins and Schrag 2001; Paul et al. 2001; Schrag et al. 1996; Schrag and DePaolo 1993; McDuff 1985) who have reconstructed LGM salinity and *δ*^{18}O from pore fluid profiles have all assumed that the bottom water salinity history was a scaled version of the benthic *δ*^{18}O_{c} history or the sea level curve. Such a scaling imposes a specific set of temporal correlations on the bottom water history. It is difficult to justify assuming that the temporal correlation of salinity at a deep ocean site exactly matches that of the changing sea level or of *δ*^{18}O_{c}. The temporal correlation in *δ*^{18}O is only equivalent to that in a benthic *δ*^{18}O_{c} record if the proportion of changes in *δ*^{18}O to temperature remains constant over time. Without this assumed temporal correlation, the uncertainty of the LGM salinity and *δ*^{18}O inferred from pore fluids is greater. If we can narrow the likely range of temporal correlation further, as is done through assumption in previous work, we can use the pore fluid profiles to help constrain the bottom water salinity and *δ*^{18}O histories from the LGM through the deglaciation to the present. Increased prior information on *T* could be gained from analyzing temporal correlations in benthic foraminiferal records of bottom water properties in the same cores.

M.D.M. thanks William Jenkins for introduction to the Lomb–Scargle periodogram and Eli Tziperman for many helpful discussions and constructive reviews of drafts of this manuscript.

## REFERENCES

Adkins, J. F., , and D. P. Schrag, 2001: Pore fluid constraints on deep ocean temperature and salinity during the last glacial maximum.

,*Geophys. Res. Lett.***28**, 771–774, doi:10.1029/2000GL011597.Adkins, J. F., , and D. P. Schrag, 2003: Reconstructing Last Glacial Maximum bottom water salinities from deep-sea sediment pore fluid profiles.

,*Earth Planet. Sci. Lett.***216**, 109–123, doi:10.1016/S0012-821X(03)00502-8.Adkins, J. F., , K. McIntyre, , and D. P. Schrag, 2002: The salinity, temperature and δ18O of the glacial deep ocean.

,*Science***298**, 1769–1773, doi:10.1126/science.1076252.Ascher, U. M., , S. J. Ruuth, , and B. T. R. Wetton, 1995: Implicit-explicit methods for time-dependent partial differential equations.

,*SIAM J. Numer. Anal.***32**, 797–823, doi:10.1137/0732037.Bard, E., , B. Hamelin, , and R. G. Fairbanks, 1990a: U-Th ages obtained by mass spectrometry in corals from Barbados: Sea level during the past 130,000 years.

,*Nature***346**, 456–458, doi:10.1038/346456a0.Bard, E., , B. Hamelin, , R. G. Fairbanks, , and A. Zindler, 1990b: Calibration of the 14C timescale over the past 30,000 years using mass spectrometric U-Th ages from Barbados corals.

,*Nature***345**, 405–410, doi:10.1038/345405a0.Bard, E., , B. Hamelin, , M. Arnold, , L. Montaggioni, , G. Cabioch, , G. Faure, , and F. Rougerie, 1996: Deglacial sea-level record from Tahiti corals and the timing of global meltwater discharge.

,*Nature***382**, 241–244, doi:10.1038/382241a0.Bellman, R. E., 1957:

*Dynamic Programming.*Princeton University Press, 384 pp.Berner, R. A., 1980:

*Early Diagenesis: A Theoretical Approach.*Princeton University Press, 241 pp.Boudreau, B. P., 1997:

*Diagenetic Models and Their Implementation: Modelling Transport and Reaction in Aquatic Sediments.*Springer-Verlag, 414 pp.Broecker, W., , and S. Barker, 2007: A 190‰ drop in atmosphere’s Δ14C during the “Mystery Interval” (17.5 to 14.5 kyr).

,*Earth Planet. Sci. Lett.***256**, 90–99, doi:10.1016/j.epsl.2007.01.015.Chappell, J., , and N. J. Shackleton, 1986: Oxygen isotopes and sea level.

,*Nature***324**, 137–140, doi:10.1038/324137a0.Chappell, J., , A. Omura, , T. Esat, , M. McCulloch, , J. Pandolfi, , Y. Ota, , and B. Pillans, 1996: Reconciliation of late Quaternary sea levels derived from coral terraces at Huon Peninsula with deep sea oxygen isotope records.

,*Earth Planet. Sci. Lett.***141**, 227–236, doi:10.1016/0012-821X(96)00062-3.Charles, C. D., , J. Lynch-Stieglitz, , U. S. Ninnemann, , and R. G. Fairbanks, 1996: Climate connections between the hemisphere revealed by deep sea sediment core/ice core correlations.

,*Earth Planet. Sci. Lett.***142**, 19–27, doi:10.1016/0012-821X(96)00083-0.Chen, J. H., , H. A. Curran, , B. White, , and G. J. Wasserburg, 1991: Precise chronology of the last interglacial period: 234U-230Th data from fossil coral reefs in the Bahamas.

,*Geol. Soc. Amer. Bull.***103**, 82–97, doi:10.1130/0016-7606(1991)103<0082:PCOTLI>2.3.CO;2.Clark, P. U., and Coauthors, 2009: The Last Glacial Maximum.

,*Science***325**, 710–714, doi:10.1126/science.1172873.Collins, L. B., , Z. R. Zhu, , K.-H. Wyrwoll, , B. G. Hatcher, , P. E. Playford, , J. H. Chen, , A. Eisenhauer, , and G. J. Wasserburg, 1993a: Late Quaternary evolution of coral reefs on a cool-water carbonate margin: The Abrolhos Carbonate Platforms, southwest Australia.

,*Mar. Geol.***110**, 203–212, doi:10.1016/0025-3227(93)90085-A.Collins, L. B., and Coauthors, 1993b: Holocene growth history of a reef complex on a cool-water carbonate margin: Easter Group of the Houtman Abrolhos, Eastern Indian Ocean.

,*Mar. Geol.***115**, 29–46, doi:10.1016/0025-3227(93)90073-5.Dansgaard, W., , and H. Tauber, 1969: Glacier oxygen-18 content and Pleistocene ocean temperatures.

,*Science***166**, 499–502, doi:10.1126/science.166.3904.499.Duplessy, J.-C., , E. Bard, , M. Arnold, , N. J. Shackleton, , J. Duprat, , and L. Labeyrie, 1991: How fast did the ocean-atmosphere system run during the last deglaciation?

,*Earth Planet. Sci. Lett.***103**, 27–40, doi:10.1016/0012-821X(91)90147-A.Duplessy, J.-C., , L. Labeyrie, , and C. Waelbroeck, 2002: Constraints on the ocean oxygen isotopic enrichment between the Last Glacial Maximum and the Holocene: Paleoceanographic implications.

,*Quat. Sci. Rev.***21**, 315–330, doi:10.1016/S0277-3791(01)00107-X.Edwards, R. L., , F. W. Taylor, , and G. J. Wasserburg, 1988: Dating earthquakes with high-precision Thorium-230 ages of very young corals.

,*Earth Planet. Sci. Lett.***90**, 371–381, doi:10.1016/0012-821X(88)90136-7.Edwards, R. L., , J. W. Beck, , G. S. Burr, , D. J. Donahue, , J. M. A. Chappell, , A. L. Bloom, , E. R. M. Druffel, , and F. W. Taylor, 1993: A large drop in atmospheric 14C/12C and reduced melting in the Younger Dryas, documented with 230Th ages of corals.

,*Science***260**, 962–968, doi:10.1126/science.260.5110.962.Emiliani, C., 1966: Paleotemperature analysis of Caribbean cores P6304-8 and P6304-9 and a generalized temperature curve for the past 425,000 years.

,*J. Geol.***74**, 109–126, doi:10.1086/627150.Esat, T. M., , M. T. McCulloch, , J. Chappell, , B. Pillans, , and A. Omura, 1999: Rapid fluctuations in sea level recorded at Huon Peninsula during the penultimate deglaciation.

,*Science***283**, 197–201, doi:10.1126/science.283.5399.197.Fairbanks, R. G., 1989: A 17,000-year glacio-eustatic sea level record: Influence of glacial melting rates on the Younger Dryas event and deep-ocean circulation.

,*Nature***342**, 637–642, doi:10.1038/342637a0.Ferrari, R., , and C. Wunsch, 2009: Ocean circulation kinetic energy: Reservoirs, sources, and sinks.

,*Annu. Rev. Fluid Mech.***41**, 253–282, doi:10.1146/annurev.fluid.40.111406.102139.Gallup, C. D., , R. L. Edwards, , and R. G. Johnson, 1994: The timing of high sea levels over the past 200,000 years.

,*Science***263**, 796–800, doi:10.1126/science.263.5148.796.Gebbie, G., 2012: Tracer transport timescales and the observed Atlantic-Pacific lag in the timing of the Last Termination.

,*Paleoceanography***27**, PA3225, doi:10.1029/2011PA002273.Gebbie, G., , and P. Huybers, 2006: Meridional circulation during the Last Glacial Maximum explored through a combination of South Atlantic

*δ*18O observations and a geostrophic inverse model.*Geochem. Geophys. Geosyst.,***7,**Q11N07, doi:10.1029/2006GC001383.Gersonde, R., and Coauthors, 1999: Southern Ocean paleoceanography.

*Proc. Ocean Drill. Program: Initial Rep.,***177,**doi:10.2973/odp.proc.ir.177.1999.Hansen, P. C., 1998:

*Rank-Deficient and Discrete Ill-Posed Problems.*Society for Industrial and Applied Mathematics, 243 pp.Hodell, D. A., , K. A. Venz, , C. D. Charles, , and U. S. Ninnemann, 2003: Pleistocene vertical carbon isotope and carbonate gradients in the South Atlantic sector of the Southern Ocean.

,*Geochem. Geophys. Geosyst.***4**, 1004, doi:10.1029/2002GC000367.Huettel, M., , and I. T. Webster, 2001: Porewater flow in permeable sediments.

*The Benthic Boundary Layer: Transport Processes and Biogeochemistry,*B. P. Boudreau and B. B. Jørgensen, Eds., Oxford University Press, 144–179.Judd, A., , and M. Hovland, 2007:

*Seabed Fluid Flow.*Cambridge University Press, 492 pp.Lado Insua, T., , A. J. Spivack, , D. Graham, , S. D’Hondt, , and K. Moran, 2014: Reconstruction of Pacific Ocean bottom water salinity during the Last Glacial Maximum.

,*Geophys. Res. Lett.***41**, 2914–2920, doi:10.1002/2014GL059575.Lamy, F., , J. Kaiser, , U. Ninnemann, , D. Hebbeln, , H. W. Arz, , and J. Stoner, 2004: Antarctic timing of surface water changes off Chile and Patagonian Ice Sheet response.

,*Science***304**, 1959–1962, doi:10.1126/science.1097863.LeGrand, P., , and C. Wunsch, 1995: Constraints from paleotracer data on the North Atlantic circulation during the Last Glacial Maximum.

,*Paleoceanography***10**, 1011–1045, doi:10.1029/95PA01455.LeGrande, A. N., , and G. A. Schmidt, 2006: Global gridded data set of the oxygen isotopic composition in seawater.

,*Geophys. Res. Lett.***33**, L12604, doi:10.1029/2006GL026011.Li, Y.-H., , and S. Gregory, 1974: Diffusion of ions in sea water and in deep-sea sediments.

,*Geochim. Cosmochim. Acta***38**, 703–714, doi:10.1016/0016-7037(74)90145-8.Ludwig, K. R., , D. R. Muhs, , K. R. Simmons, , R. B. Halley, , and E. A. Shinn, 1996: Sea-level records at ~80 ka from tectonically stable platforms: Florida and Bermuda.

,*Geology***24**, 211–214, doi:10.1130/0091-7613(1996)024<0211:SLRAKF>2.3.CO;2.Lund, D. C., , J. F. Adkins, , and R. Ferrari, 2011: Abyssal Atlantic circulation during the Last Glacial Maximum: Constraining the ratio between transport and vertical mixing.

,*Paleoceanography***26**, PA1213, doi:10.1029/2010PA001938.Lynch-Stieglitz, J., 2004: Hemispheric asynchrony of abrupt climate change.

,*Science***304**, 1919–1920, doi:10.1126/science.1100374.Malone, M. J., , J. B. Martin, , J. Schonfeld, , U. S. Ninnemann, , D. Nurnberg, , and T. S. White, 2004: The oxygen isotopic composition and temperature of Southern Ocean bottom waters during the Last Glacial Maximum.

,*Earth Planet. Sci. Lett.***222**, 275–283, doi:10.1016/j.epsl.2004.02.027.Marchal, O., , and W. B. Curry, 2008: On the abyssal circulation in the glacial Atlantic.

,*J. Phys. Oceanogr.***38**, 2014–2037, doi:10.1175/2008JPO3895.1.MARGO Project Members, 2009: Constraints on the magnitude and patterns of cooling at the Last Glacial Maximum.

,*Nat. Geosci.***2**, 127–132, doi:10.1038/ngeo411.McDuff, R. E., 1985: The chemistry of interstitial waters, Deep Sea Drilling Project Leg 86.

*Initial Reports of the Deep Sea Drilling Project,*Vol. 86, G. R. Heath et al., Eds., U.S. Government Printing Office, 675–687, doi:10.2973/dsdp.proc.86.131.1985.Miller, M. D., 2014: The deep ocean density structure at the Last Glacial Maximum: What was it and why? Ph.D. dissertation, California Institute of Technology, 230 pp.

Miller, M. D., , J. F. Adkins, , D. Menemenlis, , and M. P. Schodlok, 2012: The role of ocean cooling in setting glacial southern source bottom water salinity.

,*Paleoceanography***27**, PA3207, doi:10.1029/2012PA002297.Minson, S. E., , M. Simons, , and J. L. Beck, 2013: Bayesian inversion for finite fault earthquake source models I—Theory and algorithm.

,*Geophys. J. Int.***194**, 1701–1726, doi:10.1093/gji/ggt180.Mix, A. C., 1987: The oxygen-isotope record of glaciation.

*North America and Adjacent Oceans during the Last Deglaciation,*W. F. Ruddiman and H. E. Wright Jr., Eds., Geological Society of America, 111–135.Pahnke, K., , and R. Zahn, 2005: Southern Hemisphere water mass conversion linked with North Atlantic climate variability.

,*Science***307**, 1741–1746, doi:10.1126/science.1102163.Paul, H. A., , S. M. Bernasconi, , D. W. Schmid, , and J. A. Mckenzie, 2001: Oxygen isotopic composition of the Mediterranean Sea since the Last Glacial Maximum: Constraints from pore water analyses.

,*Earth Planet. Sci. Lett.***192**, 1–14, doi:10.1016/S0012-821X(01)00437-X.Robinson, L. F., , J. F. Adkins, , L. D. Keigwin, , J. Southon, , D. P. Fernandez, , S.-L. Wang, , and D. S. Scheirer, 2005: Radiocarbon variability in the western North Atlantic during the last deglaciation.

,*Science***310**, 1469–1473, doi:10.1126/science.1114832.Schrag, D. P., , and D. J. DePaolo, 1993: Determination of δ18O of seawater in the deep ocean during the Last Glacial Maximum.

,*Paleoceanography***8**, 1–6, doi:10.1029/92PA02796.Schrag, D. P., , G. Hampt, , and D. Murray, 1996: Pore fluid constraints on the temperature and oxygen isotopic composition of the glacial ocean.

,*Science***272**, 1930–1932, doi:10.1126/science.272.5270.1930.Schrag, D. P., , J. Adkins, , K. McIntyre, , J. Alexander, , D. Hodell, , C. Charles, , and J. McManus, 2002: The oxygen isotopic composition of seawater during the Last Glacial Maximum.

,*Quat. Sci. Rev.***21**, 331–342, doi:10.1016/S0277-3791(01)00110-X.Schmidt, G. A., , G. R. Bigg, , and E. J. Rohling, 1999: Global Seawater Oxygen-18 Database–v1.21. NASA Goddard Institute for Space Studies, accessed 12 September 2012. [Available online at http://data.giss.nasa.gov/o18data/.]

Shackleton, N. J., 1967: Oxygen isotope analyses and Pleistocene temperatures re-assessed.

,*Nature***215**, 15–17, doi:10.1038/215015a0.Skinner, L., , and N. Shackleton, 2005: An Atlantic lead over Pacific deep-water change across Termination I: Implications for the application of the marine isotope stage stratigraphy.

,*Quat. Sci. Rev.***24**, 571–580, doi:10.1016/j.quascirev.2004.11.008.Spinelli, G. A., , E. R. Giambalvo, , and A. T. Fisher, 2004: Hydrologic properties and distribution of sediments.

*Hydrogeology of the Oceanic Lithosphere,*E. Davis and H. Elderfield, Eds., Cambridge University Press, 151–188.Stein, M., , G. J. Wasserburg, , P. Aharon, , J. H. Chen, , Z. R. Zhu, , A. Bloom, , and J. Chappell, 1993: TIMS U-series dating and stable isotopes of the last interglacial event in Papua New Guinea.

,*Geochim. Cosmochim. Acta***57**, 2541–2554, doi:10.1016/0016-7037(93)90416-T.Stirling, C. H., , T. M. Esat, , M. T. McCulloch, , and K. Lambeck, 1995: High-precision U-series dating of corals from Western Australia and implications for the timing and duration of the Last Interglacial.

,*Earth Planet. Sci. Lett.***135**, 115–130, doi:10.1016/0012-821X(95)00152-3.Stirling, C. H., , T. M. Esat, , K. Lambeck, , and M. T. McCulloch, 1998: Timing and duration of the Last Interglacial: Evidence for a restricted interval of widespread coral reef growth.

,*Earth Planet. Sci. Lett.***160**, 745–762, doi:10.1016/S0012-821X(98)00125-3.Szabo, B. J., , K. R. Ludwig, , D. R. Muhs, , and K. R. Simmons, 1994: Thorium-230 ages of corals and duration of the Last Interglacial sea-level high stand on Oahu, Hawai’i.

,*Science***266**, 93–96, doi:10.1126/science.266.5182.93.Tarantola, A., 2005:

*Inverse Problem Theory and Methods for Model Parameter Estimation.*Society for Industrial and Applied Mathematics, 352 pp.Wang, J. H., , C. V. Robinson, , and I. S. Edelman, 1953: Self-diffusion and structure of liquid water. III. Measurement of the self-diffusion of liquid water with H2, H3 and O18 as tracers.

,*J. Amer. Chem. Soc.***75**, 466–470, doi:10.1021/ja01098a061.Wunsch, C., , and P. Heimbach, 2008: How long to oceanic tracer and proxy equilibrium?

,*Quat. Sci. Rev.***27**, 637–651, doi:10.1016/j.quascirev.2008.01.006.Zika, J. D., , B. M. Sloyan, , and T. J. McDougall, 2009: Diagnosing the Southern Ocean overturning from tracer fields.

,*J. Phys. Oceanogr.***39**, 2926–2940, doi:10.1175/2009JPO4052.1.Zika, J. D., , T. J. McDougall, , and B. M. Sloyan, 2010: Weak mixing in the eastern North Atlantic: An application of the tracer-contour inverse method.

,*J. Phys. Oceanogr.***40**, 1881–1893, doi:10.1175/2010JPO4360.1.

^{1}

While there is significant two- or three-dimensional advective fluid flow in many places on the ocean floor associated with spreading and converging plate boundaries, submarine groundwater discharge from continental aquifers, gas seeps (Judd and Hovland 2007), and bathymetric pressure perturbations due to current–obstruction interactions (Huettel and Webster 2001), the evolution of pore fluid concentration profiles in impermeable muddy sediments in abyssal plains is dominated by one-dimensional diffusion (Spinelli et al. 2004; Huettel and Webster 2001; Boudreau 1997; Berner 1980).